`2c722c3`

Keep quality repair worklists visible

Authored by mfwolffe <wolffemf@dukes.jmu.edu> 1 week ago

SHA: 2c722c3bb0c0301c8dff50d7e19eacc695adf94e
Parents: 432985d
Tree: 4eac6c6

2 changed files

Status	File	+	-
M	`src/loader/runtime/finalization.py`	23	5
M	`tests/test_finalization.py`	45	1

src/loader/runtime/finalization.pymodified

      Awaitable[None],
+ ]
 +_MAX_HTML_QUALITY_REPAIR_TARGETS = 8
++
  @dataclass
  class CompletionGateResult:
      targets: list[VerificationQualityRepairTarget],
  ) -> str:
      lines = ["Repair focus:"]
 -    for target in targets[:4]:
 +    visible_targets = targets[:_MAX_HTML_QUALITY_REPAIR_TARGETS]
 +    for target in visible_targets:
          lines.append(f"- Improve `{target.artifact_path}`: {target.issue}.")
 +    hidden_count = len(targets) - len(visible_targets)
 +    if hidden_count > 0:
 +        lines.append(
 +            f"- {hidden_count} additional generated artifact(s) also failed quality; "
 +            "repair the listed targets first, then rerun verification for the next batch."
 +        )
      primary = targets[0]
      lines.extend(
+         [
 -            f"- Immediate next step: edit `{primary.artifact_path}`.",
 +            f"- Immediate next step: edit `{primary.artifact_path}` with a substantial "
 +            "expansion or replacement that satisfies its listed quality issue.",
 +            "- If a target says thin content, add enough concrete prose to comfortably "
 +            "exceed the expected text-character floor; if it says insufficient "
 +            "structured content, add enough real sections, lists, code, tables, or "
 +            "other content blocks to exceed the block floor.",
 +            "- Repair every listed quality target in order before any final answer; "
 +            "do not stop after touching only the first file.",
              "- Update the listed generated artifacts directly; do not recreate the "
              "artifact set or reread unrelated reference materials.",
 -            "- After each content edit, continue with the next listed quality target "
 -            "or finish so Loader can re-run verification.",
 +            "- Once the listed target set has been substantially repaired, finish so "
 +            "Loader can rerun verification.",
+         ]
+     )
      return "\n".join(lines)
              "Do not restart discovery or keep auditing unrelated files. "
              "Your next response should be one concrete `edit` or `write`-style tool "
              f"call that expands `{primary_target.artifact_path}` to address: "
 -            f"{primary_target.issue}.{remaining_hint}"
 +            f"{primary_target.issue}. Make a substantial change that clears the stated "
 +            "threshold, not a small incremental edit, and do not summarize completion "
 +            f"after only one target is touched.{remaining_hint}"
+         )
      fixes = _extract_verification_repairs(dod.evidence, repair_targets=repair_targets)

tests/test_finalization.pymodified

+     )
      assert guidance.startswith("Repair focus:")
 -    assert f"Immediate next step: edit `{first_chapter}`." in guidance
 +    assert f"Immediate next step: edit `{first_chapter}` with a substantial" in guidance
 +    assert "Repair every listed quality target in order before any final answer" in guidance
      assert "HTML guide content quality issues" not in guidance
      assert repair is not None
      assert repair.artifact_path == str(first_chapter.resolve(strict=False))
      assert str(third_chapter.resolve(strict=False)) in repair.allowed_paths
 +def test_verification_repair_guidance_keeps_multi_file_quality_worklist(
 +    temp_dir: Path,
 +) -> None:
 +    chapters = temp_dir / "guides" / "nginx" / "chapters"
 +    chapters.mkdir(parents=True)
 +    chapter_paths = [
 +        chapters / f"{index:02d}-chapter-{index}.html"
 +        for index in range(1, 9)
 +    ]
 +    for path in chapter_paths:
 +        path.write_text(f"<h1>{path.stem}</h1>\n")
 +    dod = create_definition_of_done("Create an equally thorough HTML guide.")
 +    dod.evidence = [
 +        VerificationEvidence(
 +            command="quality",
 +            passed=False,
 +            output=(
 +                "HTML guide content quality issues:\n"
 +                + "\n".join(
 +                    f"{path}: thin content (200 text chars, expected at least 1758)"
 +                    for path in chapter_paths
 +                )
 +            ),
 +        )
 +    ]
++
 +    guidance = _build_verification_repair_guidance(
 +        dod,
 +        project_root=temp_dir,
 +    )
 +    repair = extract_active_repair_context(
 +        [Message(role=Role.USER, content=guidance)]
 +    )
++
 +    assert f"Improve `{chapter_paths[0]}`: thin content" in guidance
 +    assert f"Improve `{chapter_paths[-1]}`: thin content" in guidance
 +    assert "add enough concrete prose" in guidance
 +    assert "do not stop after touching only the first file" in guidance
 +    assert repair is not None
 +    assert repair.artifact_path == str(chapter_paths[0].resolve(strict=False))
 +    assert str(chapter_paths[-1].resolve(strict=False)) in repair.allowed_paths
++
++
  @pytest.mark.asyncio
  async def test_turn_finalizer_records_skipped_verification_observation(
      temp_dir: Path,