Port cite partial-success feedback to capabilities

CHANGELOG.md

@@ -1,6 +1,25 @@
 # Changelog
 ## [Unreleased]
 
+### Added
+
+- Native deferred Pydantic AI `RAGCapability` and `AnalysisCapability` implementations under `haiku.rag.capabilities`, with namespaced host state and lazy per-run database and sandbox resources.
+- Prior-turn RAG and analysis tool results are compacted before model requests while current-turn evidence remains intact.
+- Per-question capability request limits force a final answer from gathered evidence by removing only the exhausted capability's tools; unrelated agent and capability tools remain available.
+
+### Changed
+
+- Require `pydantic-ai-slim>=2.11,<3`; the `vertexai` optional extra now installs Pydantic AI's `google` extra.
+- The chat TUI consumes native Pydantic AI stream events. The web example uses the standard `AGUIAdapter` and emits one final state snapshot instead of forwarding sub-agent activity and per-tool state events.
+- Chat capability selection is now `haiku-rag chat --capability/-c {rag,analysis}`. Migrate from `--skill/-s`.
+- Evaluation targets are now `rag-capability` and `analysis-capability`, and the model override is `--capability-model`. Migrate from `rag-skill`, `analysis-skill`, and `--skill-model`.
+
+### Removed
+
+- The `haiku.skills` dependency, `haiku.rag.skills` modules, Python entry-point discovery, and sub-agent execution layer. Migrate `create_skill(...)` plus `SkillToolset` usage to `haiku.rag.capabilities.*.create_capability(...)` passed through `Agent(capabilities=[...])`.
+- The `haiku-rag create-skill` package generator. Compose native capabilities directly and package application-specific instructions and data in the consuming project.
+- Legacy sub-agent `ActivitySnapshotEvent` plumbing and per-tool `StateDeltaEvent` generation.
+
 ## [0.68.0] - 2026-07-24
 
 ### Added
@@ -28,22 +47,6 @@
 ### Added
 
 - `hotpotqa` evaluation dataset.
-- Native deferred Pydantic AI `RAGCapability` and `AnalysisCapability` implementations under `haiku.rag.capabilities`, with namespaced host state and lazy per-run database and sandbox resources.
-- Prior-turn RAG and analysis tool results are compacted before model requests while current-turn evidence remains intact.
-- Per-question capability request limits force a final answer from gathered evidence by removing only the exhausted capability's tools; unrelated agent and capability tools remain available.
-
-### Changed
-
-- Require `pydantic-ai-slim>=2.11,<3`; the `vertexai` optional extra now installs Pydantic AI's `google` extra.
-- The chat TUI consumes native Pydantic AI stream events. The web example uses the standard `AGUIAdapter` and emits one final state snapshot instead of forwarding sub-agent activity and per-tool state events.
-- Chat capability selection is now `haiku-rag chat --capability/-c {rag,analysis}`. Migrate from `--skill/-s`.
-- Evaluation targets are now `rag-capability` and `analysis-capability`, and the model override is `--capability-model`. Migrate from `rag-skill`, `analysis-skill`, and `--skill-model`.
-
-### Removed
-
-- The `haiku.skills` dependency, `haiku.rag.skills` modules, Python entry-point discovery, and sub-agent execution layer. Migrate `create_skill(...)` plus `SkillToolset` usage to `haiku.rag.capabilities.*.create_capability(...)` passed through `Agent(capabilities=[...])`.
-- The `haiku-rag create-skill` package generator. Compose native capabilities directly and package application-specific instructions and data in the consuming project.
-- Legacy sub-agent `ActivitySnapshotEvent` plumbing and per-tool `StateDeltaEvent` generation.
 
 ### Changed
 

haiku_rag_slim/haiku/rag/capabilities/_base.py

@@ -264,6 +264,15 @@ class RAGCapabilityBase[StateT: BaseModel](AbstractCapability[Any]):
                 "Copy chunk_ids verbatim from search results."
             )
         self._register_citations(citations)
+        resolved = {citation.chunk_id for citation in citations}
+        unresolved = [cid for cid in missing if cid not in resolved]
+        if unresolved:
+            return (
+                f"Registered {len(citations)} citation(s); "
+                f"ignored {len(unresolved)} unresolvable id(s): "
+                f"{unresolved}. Copy chunk_ids verbatim from search "
+                "results and cite again."
+            )
         return f"Registered {len(citations)} citation(s)."
 
     def _register_citations(self, citations: list[Citation]) -> None:

tests/capabilities/test_capabilities.py

@@ -326,6 +326,32 @@ async def test_cite_resolves_direct_chunk_ids_and_reuses_document_lookup(temp_db
     client.get_document_by_id.assert_awaited_once_with("doc-1")
 
 
+@pytest.mark.asyncio
+async def test_cite_reports_unresolved_ids_on_partial_success(temp_db_path):
+    capability = create_rag(db_path=temp_db_path, config=AppConfig())
+    capability.state = RAGState()
+    client = AsyncMock()
+    client.get_chunk_by_id.side_effect = [
+        Chunk(id="chunk-1", document_id="doc-1", content="first"),
+        None,
+        None,
+    ]
+    client.get_document_by_id.return_value = SimpleNamespace(
+        uri="test://document",
+        title="Document",
+        metadata={},
+    )
+    capability.rag = client
+
+    result = await capability._cite(["chunk-1", "6.43", "6.51.2"])
+
+    assert "Registered 1 citation(s)" in result
+    assert "6.43" in result
+    assert "6.51.2" in result
+    assert "verbatim" in result
+    assert capability.state.citations == ["chunk-1"]
+
+
 @pytest.mark.asyncio
 async def test_analysis_records_new_sandbox_search_results(temp_db_path):
     capability = create_analysis(db_path=temp_db_path, config=AppConfig())