fix(polish-summaries): Batch API custom_id charset (hash ids, path in feature) (#91)

* fix(polish-summaries): Batch API custom_id charset — hash ids, path in feature

First live run 400'd: custom_id must match ^[a-zA-Z0-9_-]{1,64}$ and
'sum__cli/concept.md' doesn't. custom_id becomes sum_<sha256[:24]>;
the readable rel path rides in the request/result 'feature' field and
apply_results maps through that instead of parsing the id. Regression
test pins the charset.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* style(tests): pinned-black formatting for test_anthropic_fable.py

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Author: silversurfer562

src/attune_author/summaries_polish.py

@@ -131,8 +131,16 @@ def polish_model() -> str:
 
 
 def custom_id_for(rel_path: str) -> str:
-    """Stable Anthropic ``custom_id`` for one template path."""
-    return f"sum__{rel_path}"
+    """Stable Anthropic ``custom_id`` for one template path.
+
+    The Batch API requires ``^[a-zA-Z0-9_-]{1,64}$`` — template paths
+    (slashes, dots) violate it, so the id is a path hash. The readable
+    path rides in the request's ``feature`` field, which ``poll_batch``
+    copies onto each result; ``apply_results`` maps results back
+    through THAT, never by parsing the id. (Learned from a live 400 on
+    the first attune-ai run, 2026-07-10.)
+    """
+    return f"sum_{_sha256(rel_path)[:24]}"
 
 
 def build_requests(help_dir: Path, model: str, *, force: bool = False) -> BuildReport:
@@ -172,8 +180,10 @@ def build_requests(help_dir: Path, model: str, *, force: bool = False) -> BuildR
         )
         requests.append(
             BatchPolishRequest(
+                # feature carries the FULL relative path — apply_results
+                # maps batch results back through it (see custom_id_for).
                 custom_id=custom_id_for(rel_path),
-                feature=str(Path(rel_path).parent),
+                feature=rel_path,
                 depth=Path(rel_path).stem,
                 system=_SYSTEM,
                 user_message=user_message,
@@ -232,7 +242,7 @@ def apply_results(
     stamp = (_now or (lambda: datetime.now(timezone.utc)))().isoformat()
     counters = {"applied": 0, "truncated": 0, "errored": 0, "empty": 0}
     for result in results:
-        rel_path = result.custom_id.removeprefix("sum__")
+        rel_path = result.feature  # full relative path — see custom_id_for
         if result.error is not None or result.text is None:
             counters["errored"] += 1
             continue

tests/test_anthropic_fable.py

@@ -137,9 +137,7 @@ def test_fable_response_leading_thinking_block_skipped() -> None:
     is the answer. Hit live 2026-07-10 via the commit regen hook."""
     thinking = SimpleNamespace(thinking="reasoning...", type="thinking")
     text_block = SimpleNamespace(text="the answer", type="text")
-    response = SimpleNamespace(
-        content=[thinking, text_block], stop_reason="end_turn", usage=None
-    )
+    response = SimpleNamespace(content=[thinking, text_block], stop_reason="end_turn", usage=None)
     client = MagicMock()
     client.beta.messages.create.return_value = response
     assert (

tests/test_summaries_polish.py

@@ -40,7 +40,8 @@ def test_builds_one_request_per_entry(self, tmp_path):
         report = sp.build_requests(help_dir, "claude-sonnet-5")
         assert len(report.requests) == 2
         req = report.requests[0]
-        assert req.custom_id == "sum__cli/concept.md"
+        assert req.custom_id == sp.custom_id_for("cli/concept.md")
+        assert req.feature == "cli/concept.md"  # apply maps results via this
         assert req.model == "claude-sonnet-5"
         assert "Current (mechanical) summary: seed" in req.user_message
         assert "Body prose here." in req.user_message
@@ -78,7 +79,7 @@ def test_orphaned_sidecar_key_ignored(self, tmp_path):
         sidecar["gone/nope.md"] = "orphan"
         (help_dir / "summaries.json").write_text(json.dumps(sidecar))
         report = sp.build_requests(help_dir, "claude-sonnet-5")
-        assert [r.custom_id for r in report.requests] == ["sum__cli/concept.md"]
+        assert [r.feature for r in report.requests] == ["cli/concept.md"]
 
 
 class TestModelAndCost:
@@ -126,8 +127,8 @@ def test_strips_wrapping_quotes(self):
 class TestApplyResults:
     def _result(self, rel: str, text: str | None, error: str | None = None) -> BatchPolishResult:
         return BatchPolishResult(
-            custom_id=f"sum__{rel}",
-            feature=str(Path(rel).parent),
+            custom_id=sp.custom_id_for(rel),
+            feature=rel,
             depth=Path(rel).stem,
             text=text,
             error=error,
@@ -286,8 +287,8 @@ def test_source_hash_falls_back_to_body_sha(self, tmp_path):
             help_dir,
             [
                 BatchPolishResult(
-                    custom_id="sum__cli/concept.md",
-                    feature="cli",
+                    custom_id=sp.custom_id_for("cli/concept.md"),
+                    feature="cli/concept.md",
                     depth="concept",
                     text="Polished.",
                     error=None,
@@ -306,8 +307,8 @@ def test_apply_truncates_overlong_result_and_flags_meta(self, tmp_path):
             help_dir,
             [
                 BatchPolishResult(
-                    custom_id="sum__cli/concept.md",
-                    feature="cli",
+                    custom_id=sp.custom_id_for("cli/concept.md"),
+                    feature="cli/concept.md",
                     depth="concept",
                     text=long_text,
                     error=None,
@@ -358,3 +359,17 @@ def fake_poll(batch_id, expected, **kwargs):
         assert sidecar["cli/concept.md"] == "Resumed polish."
         assert not (help_dir / ".summaries-batch-state.json").exists()
         assert "applied 1" in capsys.readouterr().out
+
+
+class TestCustomIdContract:
+    def test_matches_batch_api_charset(self):
+        """The Batch API enforces ^[a-zA-Z0-9_-]{1,64}$ — a raw template
+        path (slashes, dots) 400s the whole submit. Hit live 2026-07-10."""
+        import re
+
+        for rel in ("cli/concept.md", "elicitation-forms/troubleshooting.md"):
+            cid = sp.custom_id_for(rel)
+            assert re.fullmatch(r"[a-zA-Z0-9_-]{1,64}", cid), cid
+
+    def test_distinct_paths_get_distinct_ids(self):
+        assert sp.custom_id_for("a/b.md") != sp.custom_id_for("a/c.md")