Merge remote-tracking branch 'origin/main' into fix/403-cached-tracked-runs

Since #347 a cached re-run of runtime-linked prompt text (PreviewAny or an
LLM into a PromptManager text input) counts in the Library, but its history
capture had no raw text, so the history page's "tracked runs" stayed at the
number of jobs in which the node executed.

The #347 receipt now also keeps the raw, prepend and append text the node
gave the history capture. When a cached node's graph inference leaves a field
unknown, the capture takes it from the receipt that matches the cache entry,
labeled inferred/graph_semantics, so the job joins the prompt's history scope
and gets its prompt hint. No schema change and no new evidence vocabulary;
no backfill. Fails closed on no receipt, a disagreeing known field, equal
outputs with different fields, oversized or invalid text and a full budget
(least recently used receipts give up their fields first; counting stays).

TESTING.md

@@ -5572,8 +5572,26 @@ decide through the same `own_counts`.
 - A cached node counts at job start like the hook: a job failing later still counts it.
   A node that executes after its replay in the same job (RAM-pressure freed the entry)
   skips its own count but still consumes the pending first-use catch-up.
-- No backfill: earlier cached jobs keep their counts. The history page's "tracked runs"
-  can stay below the Library count for these prompts (#403).
+- No backfill: earlier cached jobs keep their counts.
+- History (#403): the receipt also keeps the raw, prepend and append text the node gave
+  the history capture. For a cached node whose graph inference leaves a field unknown,
+  the capture takes it from the receipt matching the cache entry
+  (`CachedRuns.recorded_fields`), as `inferred` / `graph_semantics`: the node did not run
+  in this job, the value follows from the run key (the whole upstream closure) plus the
+  cache proof. The job then joins the prompt's history scope, gets the
+  `positive_prompt_id` (and Pair) hint, and an image of it can be attributed through its
+  run, so the history page's "tracked runs" equals the Library count from the restart on.
+  Fail closed (fields stay unknown, as before): no matching receipt, a receipt that
+  disagrees with a field the graph knows, executions with equal outputs that recorded
+  different fields (a conflict that stays for that run key), a field over 64 KiB or
+  invalid Unicode, more than one run's history (1 MiB) in one receipt, or a full 8 MiB
+  field budget (least recently used receipts give up their fields first; their counting
+  stays). No new evidence vocabulary, so older builds still open the database.
+- Still different: jobs before the restart (no backfill); list batches (per item in the
+  Library, once per job in history); text with leading or trailing whitespace (the
+  Library stores it stripped, history matches the exact raw text, #438); a job in which the
+  node executes after its replay (RAM-pressure freed the entry: two samples, an
+  ambiguous capture); jobs while the history writer is fenced (restore, library move).
 - Tests: `tests/test_cached_runs.py` (whole jobs with the real nodes, a real database,
   fake host caches); `tests/comfyui/execution-text-link.spec.mjs` "#347" cases on the
   managed host: PreviewAny into each node kind, 3 jobs changing only the sampler seed,
@@ -5581,7 +5599,10 @@ decide through the same `own_counts`.
   modes, A, B, A from the owned `PMCycledText` under LRU, typed text, a downstream
   failure, a "run selected output" job that does not need the node, and a rewire to the
   negative input. The RAM-pressure eviction path cannot be forced on a real host and is
-  covered by the Python tests.
+  covered by the Python tests. Since #403 these cases also check the history page's
+  tracked runs equal the Library count, and `tests/test_execution_history_hooks.py`
+  (`CachedRuntimeHistoryTests`) runs the real history hook and writer with a real
+  `CachedRuns`.
 
 ```sh
 python3 -m unittest discover -s tests -p "test_cached_runs.py"
@@ -5680,10 +5701,12 @@ Context cases verify A,A,B as three jobs/two combinations/one positive snapshot
 when every positive field matches; all five registered PM variants; exact raw
 whitespace and Unicode; known empty negatives; Text negative roles and crossed
 pairs; fully cached repeats after trigger configuration changes; successful source
-text before downstream failure/interruption; unknown cached dynamic raw; unsupported
+text before downstream failure/interruption; cached dynamic raw from the reused execution (#403); unsupported
 transforms and duplicate list invocation ambiguity. Fresh values use
 `runtime_return`, cached effective STRING outputs use `host_cache_output`, and
-submitted literals use graph inference. SQL NULL means unknown; empty text is
+submitted literals use graph inference. A cached dynamic raw is unknown unless the
+#347 receipt of the execution the cache entry comes from proves it; it is then
+`inferred` / `graph_semantics` (graph identity plus a same-process cache receipt, #403). SQL NULL means unknown; empty text is
 known. Text is never trimmed or normalized by capture. Cache availability and
 inferred sampler roles prove neither sampler execution nor consumption, and list
 indices do not identify generated images. External trigger configuration changes

prompt_manager.py

@@ -275,7 +275,12 @@ class PromptManager(PromptManagerBase, ComfyNodeABC):
         record_runtime_context(
             type(self).__name__, text, prepend_text, append_text, encoding_text
         )
-        self._record_cached_item(unique_id, (encoding_text,), storage_text)
+        self._record_cached_item(
+            unique_id,
+            (encoding_text,),
+            storage_text,
+            fields=((text, prepend_text, append_text),),
+        )
         return result
 
     @classmethod

prompt_manager_base.py

@@ -243,14 +243,24 @@ class PromptManagerBase:
             return False
 
     def _record_cached_item(
-        self, unique_id: Any, outputs: tuple, text: Any, negative_text: Any = None
+        self,
+        unique_id: Any,
+        outputs: tuple,
+        text: Any,
+        negative_text: Any = None,
+        fields: Optional[tuple] = None,
     ) -> None:
-        """Remember this execution's outputs and stored text for cached jobs (#347)."""
+        """Remember this execution's outputs and stored text for cached jobs (#347).
+
+        ``fields`` are the ``(raw, prepend, append)`` texts this execution gave
+        ``record_runtime_context`` for each output, so a cached job's history
+        can record them too (#403).
+        """
         if unique_id is None:
             return
         try:
             self._cached_runs().record_item(
-                unique_id, outputs, stored_facts(text, negative_text)
+                unique_id, outputs, stored_facts(text, negative_text), fields
             )
         except Exception as error:
             self.logger.debug(f"Could not record the cached-run receipt: {error}")

prompt_manager_pair.py

@@ -198,7 +198,11 @@ class PromptManagerPair(PromptManagerFocused):
             type(self).__name__, negative_text, "", "", negative_text, role="negative"
         )
         self._record_cached_item(
-            unique_id, (positive, negative_text), text, negative_text
+            unique_id,
+            (positive, negative_text),
+            text,
+            negative_text,
+            fields=((text, prepend_text, append_text), (negative_text, "", "")),
         )
         return result
 

prompt_manager_text.py

@@ -246,7 +246,12 @@ class PromptManagerText(PromptManagerBase, ComfyNodeABC):
         record_runtime_context(
             type(self).__name__, text, prepend_text, append_text, final_text
         )
-        self._record_cached_item(unique_id, (final_text,), storage_text)
+        self._record_cached_item(
+            unique_id,
+            (final_text,),
+            storage_text,
+            fields=((text, prepend_text, append_text),),
+        )
         return result
 
     @classmethod

tests/comfyui/execution-history.spec.mjs

@@ -263,15 +263,18 @@ test.describe('public chronological history with the actual acknowledged host br
     });
 });
 
-test('a cached dynamic unknown positive stays readable by run UUID and outside the known raw Library scope', async ({ execution }) => {
-    const text = `unknown positive ${randomUUID()}`, graph = recipeGraph({ text, dynamic: true });
+// #403: the graph cannot show a dynamic positive's raw text; a cached job takes it from the execution it reuses.
+test('a cached dynamic positive takes its raw text from the execution it reuses and joins the Library scope', async ({ execution }) => {
+    const text = `dynamic positive ${randomUUID()}`, graph = recipeGraph({ text, dynamic: true });
     const first = await submit(execution.request, graph);await history(execution.request, first);const known = (await capture(execution, first)).contexts[0];
-    const second = await submit(execution.request, graph);await history(execution.request, second);const unknown = (await capture(execution, second)).contexts[0];
-    expect(unknown.positive.raw).toBeNull();expect(unknown.positive.effective).toBe(known.positive.effective);
+    const second = await submit(execution.request, graph);await history(execution.request, second);const reused = await capture(execution, second);const cached = reused.contexts[0];
+    expect(cached.evidence.positive.effective.source).toBe('host_cache_output');expect(cached.positive.effective).toBe(known.positive.effective);
+    expect(cached.positive.raw).toBe(known.positive.raw);expect(cached.evidence.positive.raw).toMatchObject({ provenance: 'inferred', source: 'graph_semantics' });
+    expect(cached.positive_prompt_id).toBe(known.positive_prompt_id);
     const result = await readExecutionRuns(execution.databasePath, second), runId = result.rows[0].run_id;
-    const detail = await execution.request.get(`/prompt_manager/history/runs/${runId}`);expect(detail.ok()).toBe(true);expect((await detail.json()).contexts[0].positive.raw).toBeNull();
+    const detail = await execution.request.get(`/prompt_manager/history/runs/${runId}`);expect(detail.ok()).toBe(true);expect((await detail.json()).contexts[0].positive.raw).toBe(known.positive.raw);
     const response = await execution.request.get(`/prompt_manager/library/${known.positive_prompt_id}/history`);expect(response.ok()).toBe(true);const window = await response.json();
-    expect(window.recorded_run_count).toBe(1);expect(window.items.map(item => item.run_id)).not.toContain(runId);
+    expect(window.recorded_run_count).toBe(2);expect(window.items.map(item => item.run_id)).toContain(runId);
 });
 
 test('actual PM A,A,B jobs persist two combinations and one identical positive snapshot', async ({ execution }) => {
@@ -369,7 +372,7 @@ test('successful source capture survives downstream failure and real interruptio
     }
 });
 
-test('cached dynamic raw stays unknown while literal affixes and effective cache text survive', async ({ execution }) => {
+test('cached dynamic raw comes from the reused execution while literal affixes and effective cache text survive (#403)', async ({ execution }) => {
     const graph = recipeGraph({ dynamic: true, text: 'runtime dynamic recipe', prepend: ' pre ', append: ' post ' });
     const first = await submit(execution.request, graph);
     await history(execution.request, first);
@@ -377,8 +380,9 @@ test('cached dynamic raw stays unknown while literal affixes and effective cache
     const second = await submit(execution.request, graph);
     await history(execution.request, second);
     const value = (await capture(execution, second)).contexts[0];
-    expect(value.positive).toEqual({ raw: null, prepend: ' pre ', append: ' post ', effective: 'pre runtime dynamic recipe post' });
-    expect(value.evidence.positive.raw.provenance).toBe('unknown');
+    expect(value.positive).toEqual({ raw: 'runtime dynamic recipe', prepend: ' pre ', append: ' post ', effective: 'pre runtime dynamic recipe post' });
+    expect(value.evidence.positive.raw).toMatchObject({ provenance: 'inferred', source: 'graph_semantics' });
+    expect(value.evidence.positive.prepend.source).toBe('graph_literal');
     expect(value.evidence.positive.effective.source).toBe('host_cache_output');
 });
 

tests/comfyui/execution-text-link.spec.mjs

@@ -287,6 +287,14 @@ async function capturedContext(execution, id) {
     return result.captures[result.rows[0].run_id].contexts[0];
 }
 
+// #403: the prompt's history "tracked runs", once every job's capture reached SQLite.
+async function expectTrackedRuns(execution, promptId, expected) {
+    await expect.poll(async () => {
+        const response = await execution.request.get(`/prompt_manager/library/${promptId}/history`);
+        return response.ok() ? (await response.json()).recorded_run_count : null;
+    }, { timeout: 10000, message: 'history tracked runs equal the Library run count' }).toBe(expected);
+}
+
 async function firstRun(execution, page, testCase) {
     const prefix = `textlink/${randomUUID()}`;
     const built = await build(page, { ...testCase, prefix });
@@ -309,6 +317,8 @@ for (const kind of ['PromptManager', 'PromptManagerFocused', 'PromptManagerSplit
         const counts = await runCounts(execution, first.context);
         expect(counts.positive).toBe(3);
         if (pair) expect(counts).toMatchObject({ negative: 0, pair: 3 });
+        // #403: the cached jobs are tracked runs of the prompt too
+        await expectTrackedRuns(execution, first.context.positive_prompt_id, 3);
     });
 }
 
@@ -329,6 +339,7 @@ test.describe('cached runs in one renderer (#347)', () => {
                     else expect(again.cached).toContain(first.promptNode);
                 }
                 expect((await runCounts(execution, first.context)).positive).toBe(3);
+                await expectTrackedRuns(execution, first.context.positive_prompt_id, 3);
             } finally { await setCache(execution, 'RAM_PRESSURE'); }
         });
     }
@@ -352,6 +363,9 @@ test.describe('cached runs in one renderer (#347)', () => {
             const bContext = await capturedContext(execution, second.id);
             expect(bContext.positive.raw).toBe(b);
             expect((await runCounts(execution, bContext)).positive).toBe(1);
+            // #403: the third job reuses A's execution, not the newest (B's)
+            await expectTrackedRuns(execution, first.context.positive_prompt_id, 2);
+            await expectTrackedRuns(execution, bContext.positive_prompt_id, 1);
         } finally { await setCache(execution, 'RAM_PRESSURE'); }
     });
 
@@ -390,10 +404,14 @@ test.describe('cached runs in one renderer (#347)', () => {
         } });
         expect(rewired.cached).toContain(first.promptNode);
         expect((await runCounts(execution, first.context)).positive).toBe(1);
-        // The history capture of a cached job is unchanged: effective text from the cache, no stored-prompt hint
+        // #403: a cached job's history takes the raw text of the execution it reuses and names the prompt
         const cachedContext = await capturedContext(execution, (await queueAgain(execution, first, { seed: 3 })).id);
         expect(cachedContext.positive.effective).toBe(first.context.positive.effective);
-        expect(cachedContext.positive_prompt_id).toBeNull();
+        expect(cachedContext.positive.raw).toBe(first.context.positive.raw);
+        expect(cachedContext.evidence.positive.raw).toMatchObject({ provenance: 'inferred', source: 'graph_semantics' });
+        expect(cachedContext.positive_prompt_id).toBe(first.context.positive_prompt_id);
         expect((await runCounts(execution, first.context)).positive).toBe(2);
+        // The rewired job used the prompt as its negative: neither number counts it
+        await expectTrackedRuns(execution, first.context.positive_prompt_id, 2);
     });
 });

tests/test_cached_runs.py

@@ -666,7 +666,7 @@ class TestOwnCounts(unittest.TestCase):
         return counts
 
 
-class TestReceipts(unittest.TestCase):
+class ReceiptCase(unittest.TestCase):
     def setUp(self):
         self.context = ("j1", None)
         self.runs = CachedRuns(
@@ -686,6 +686,8 @@ class TestReceipts(unittest.TestCase):
     def keys(self):
         return list(self.runs._receipts)
 
+
+class TestReceipts(ReceiptCase):
     def test_least_recently_used_run_keys_are_dropped(self):
         self.begin(1)
         for name in ("a", "b", "c"):
@@ -726,6 +728,196 @@ class TestReceipts(unittest.TestCase):
         self.assertEqual(list(self.runs._attempt.nodes), ["b"])
 
 
+class TestRecordedFields(ReceiptCase):
+    """The texts a receipt keeps for the history capture of a cached job (#403)."""
+
+    F = (("raw a", " pre ", ""),)
+
+    def record(self, item_id, name="a", output="x", fields=F, facts=None):
+        self.begin(item_id)
+        self.runs.record_item(name, (output,), facts or Facts("h"), fields)
+        self.runs.end_attempt(item_id)
+
+    def lookup(self, name="a", values=(("x",),), prompt_id="j9", item_id=9):
+        self.begin(item_id, prompt_id)
+        return self.runs.recorded_fields(prompt_id, name, values)
+
+    def test_a_cached_job_reads_the_fields_of_the_execution_it_reuses(self):
+        self.record(1)
+        self.assertEqual(self.lookup(), [self.F])
+
+    def test_list_items_return_their_own_fields_in_order(self):
+        self.begin(1)
+        for index, text in enumerate(("x", "y")):
+            self.context = ("j1", index)
+            self.runs.record_item("a", (text,), Facts("h"), ((text, "", ""),))
+        self.runs.end_attempt(1)
+        self.assertEqual(
+            self.lookup(values=(("x", "y"),)), [(("x", "", ""),), (("y", "", ""),)]
+        )
+
+    def test_fails_closed(self):
+        self.record(1)
+        for case, kwargs in (
+            ("other output", dict(values=(("changed",),))),
+            ("other node", dict(name="b")),
+            ("unknown node", dict(name="zzz")),
+            ("no values", dict(values=())),
+            ("malformed values", dict(values="x")),
+        ):
+            with self.subTest(case=case):
+                self.assertIsNone(self.lookup(**kwargs))
+        with self.subTest(case="another job"):
+            self.begin(9, "j9")
+            self.assertIsNone(self.runs.recorded_fields("j8", "a", (("x",),)))
+        with self.subTest(case="no job"):
+            self.runs.end_attempt(9)
+            self.assertIsNone(self.runs.recorded_fields("j9", "a", (("x",),)))
+
+    def test_an_execution_without_fields_gives_none(self):
+        for fields in (None, (("a", "b"),), ((1, "", ""),), (), "abc"):
+            with self.subTest(fields=fields):
+                self.runs = CachedRuns(context=lambda: self.context)
+                self.record(1, fields=fields)
+                self.assertIsNone(self.lookup())
+
+    def test_a_lookup_never_counts_or_marks_the_node_replayed(self):
+        self.record(1)
+        self.lookup()
+        self.assertFalse(self.runs.replayed("a"))
+
+    def test_equal_outputs_with_other_raw_text_conflict_and_still_count(self):
+        key = prompt_node_key(self.g, "a")
+        self.record(1, fields=(("x ", "", ""),))
+        self.record(2, fields=(("x", "", ""),))
+        # Counting dedupes by outputs as before; the fields cannot name either
+        self.assertEqual([r.item_id for r in self.runs._receipts[key]], [2])
+        self.assertIsNone(self.lookup())
+        self.assertIsNotNone(self.runs._match(key, (("x",),)))
+        # The conflict stays: a later execution cannot tell which text it was
+        self.record(3, fields=(("x", "", ""),))
+        self.assertIsNone(self.lookup())
+
+    def test_equal_outputs_with_other_facts_still_count_the_newest(self):
+        key = prompt_node_key(self.g, "a")
+        self.record(1, facts=Facts("old"), fields=(("a x", "", ""),))
+        self.record(2, facts=Facts("new"), fields=(("x", "a ", ""),))
+        items = self.runs._match(key, (("x",),))
+        self.assertEqual([item.facts for item in items], [Facts("new")])
+
+    def test_equal_outputs_and_fields_still_replace(self):
+        key = prompt_node_key(self.g, "a")
+        self.record(1)
+        self.record(2)
+        self.assertEqual([r.item_id for r in self.runs._receipts[key]], [2])
+        self.assertEqual(self.lookup(), [self.F])
+
+    def test_fields_dropped_by_the_budget_do_not_conflict(self):
+        self.runs = CachedRuns(context=lambda: self.context, max_field_bytes=0)
+        self.record(1)
+        self.runs._max_field_bytes = 100
+        self.record(2)
+        self.assertEqual(self.lookup(), [self.F])
+
+    def test_oversized_or_invalid_fields_are_not_kept(self):
+        for raw in ("x" * (module.MAX_FIELD_BYTES + 1), "\ud800"):
+            with self.subTest(size=len(raw)):
+                self.runs = CachedRuns(context=lambda: self.context)
+                self.record(1, fields=((raw, "", ""),))
+                self.assertIsNone(self.lookup())
+                self.assertEqual(self.runs._field_bytes, 0)
+
+    def test_old_receipts_give_up_their_fields_to_new_ones(self):
+        self.runs = CachedRuns(
+            context=lambda: self.context, max_keys=3, max_field_bytes=10
+        )
+        self.record(1, "a", fields=(("123456", "", ""),))
+        self.record(2, "b", fields=(("123456", "", ""),))
+        self.assertEqual(self.runs._field_bytes, 6)
+        self.assertIsNone(self.lookup("a"))
+        self.assertEqual(self.lookup("b"), [(("123456", "", ""),)])
+        # "a" still counts
+        self.assertIsNotNone(self.runs._match(prompt_node_key(self.g, "a"), (("x",),)))
+        # Larger than the whole budget: not kept, nothing else given up
+        self.record(3, "c", fields=(("12345678901", "", ""),))
+        self.assertEqual(self.runs._field_bytes, 6)
+        self.assertEqual(self.lookup("b"), [(("123456", "", ""),)])
+
+    def test_one_receipt_holds_at_most_one_runs_history(self):
+        with mock.patch.object(module, "MAX_RECEIPT_FIELD_BYTES", 8):
+            self.begin(1)
+            for index, raw in enumerate(("12345", "12345")):
+                self.context = ("j1", index)
+                self.runs.record_item("a", (raw,), Facts("h"), ((raw, "", ""),))
+            self.runs.end_attempt(1)
+        self.assertEqual(self.runs._field_bytes, 5)
+        self.assertIsNone(self.lookup(values=(("12345", "12345"),)))
+
+    def test_the_byte_count_follows_replacement_and_eviction(self):
+        self.begin(1)
+        self.runs.record_item("a", ("x",), Facts("h"), (("12345", "", ""),))
+        self.runs.record_item("a", ("x",), Facts("h"), (("123", "", ""),))
+        self.assertEqual(self.runs._field_bytes, 3)
+        self.runs.end_attempt(1)
+        for item_id, raw in ((2, "1"), (3, "22"), (4, "333")):
+            self.record(item_id, output=raw, fields=((raw, "", ""),))
+        # max_receipts=2: only the receipts of jobs 4 and 3 remain
+        self.assertEqual(self.runs._field_bytes, 5)
+        for item_id, name in ((5, "b"), (6, "c")):  # max_keys=2: "a" is dropped
+            self.record(item_id, name, fields=(("1", "", ""),))
+        self.assertEqual(self.runs._field_bytes, 2)
+        self.assertEqual(
+            self.runs._field_bytes,
+            sum(module._bytes(r) for r in self.runs._receipts.values()),
+        )
+
+
+class TestCachedJobsRecordFields(JobCase):
+    """The real nodes keep the texts they gave the history capture (#403)."""
+
+    def lookup(self, g, node_id="pm", prompt_id="job-x"):
+        self.item_id += 1
+        self.runs.begin_attempt(
+            ((self.item_id, prompt_id, g, {}, ["save"], {}), self.item_id)
+        )
+        values = cached_string_outputs(self.executor, node_id, g[node_id]["class_type"])
+        found = self.runs.recorded_fields(prompt_id, node_id, values)
+        self.runs.end_attempt(self.item_id)
+        return found
+
+    def test_each_node_kind_keeps_its_runtime_fields(self):
+        for kind in (
+            "PromptManager",
+            "PromptManagerFocused",
+            "PromptManagerSplit",
+            "PromptManagerText",
+        ):
+            with self.subTest(kind=kind):
+                g = graph(kind)
+                self.first(g, text=f" {LLM} ")
+                self.assertEqual(
+                    self.lookup(graph(kind, seed=2)), [((f" {LLM} ", "", ""),)]
+                )
+
+    def test_pair_keeps_positive_and_negative_fields(self):
+        g = graph("PromptManagerPair", negative_text="typed negative")
+        self.first(g, negative="typed negative")
+        self.assertEqual(
+            self.lookup(
+                graph("PromptManagerPair", negative_text="typed negative", seed=2)
+            ),
+            [((LLM, "", ""), ("typed negative", "", ""))],
+        )
+
+    def test_a_disabled_node_keeps_no_fields(self):
+        g = graph()
+        g["pm"]["inputs"]["save_to_library"] = False
+        self.first(g)
+        g2 = graph(seed=2)
+        g2["pm"]["inputs"]["save_to_library"] = False
+        self.assertIsNone(self.lookup(g2))
+
+
 class TestCacheReader(unittest.TestCase):
     def setUp(self):
         patcher = mock.patch.dict(sys.modules, host_modules())

tests/test_execution_context_host.py

@@ -100,14 +100,19 @@ class HostTests(unittest.TestCase):
         self.executor.caches = types.SimpleNamespace(outputs=HierarchicalCache())
 
     def begin(
-        self, kind="PromptManager", raw=" raw ", prepend=" pre ", append=" post "
+        self,
+        kind="PromptManager",
+        raw=" raw ",
+        prepend=" pre ",
+        append=" post ",
+        negative="negative",
     ):
         g = {
             "p": {
                 "class_type": kind,
                 "inputs": {
                     "text": raw,
-                    "negative_text": "negative",
+                    "negative_text": negative,
                     "prepend_text": prepend,
                     "append_text": append,
                 },
@@ -125,7 +130,7 @@ class HostTests(unittest.TestCase):
                 },
             },
         }
-        if type(raw) is list or type(prepend) is list:
+        if any(type(value) is list for value in (raw, prepend, negative)):
             g["dynamic"] = {"class_type": "RuntimeGenerator", "inputs": {}}
             g["stateful"] = {
                 "class_type": "PromptManagerText",
@@ -445,3 +450,138 @@ class HostTests(unittest.TestCase):
         ref = install_and_remove()
         gc.collect()
         self.assertIsNone(ref())
+
+    # -- #403: a cached runtime text takes its fields from the reused execution
+
+    class Receipts:
+        """A fake ``CachedRuns.recorded_fields``: fixed fields, calls recorded."""
+
+        def __init__(self, fields):
+            self.fields, self.calls = fields, []
+
+        def recorded_fields(self, prompt_id, node_id, values):
+            self.calls.append((prompt_id, node_id, values))
+            return self.fields
+
+    def observe_with(self, receipts, outputs, **begin):
+        self.begin(**begin)
+        self.executor.caches.outputs.entries["p"] = Entry(None, outputs)
+        self.observe("execution_start")
+        host.observe_cached_context(
+            self.executor,
+            "execution_cached",
+            {"prompt_id": JOB, "nodes": ["p"]},
+            receipts=receipts,
+        )
+        return self.seal()
+
+    def test_cached_runtime_raw_comes_from_the_matching_receipt(self):
+        receipts = self.Receipts([(("LLM raw", " pre ", " post "),)])
+        result = self.observe_with(
+            receipts, [object(), ["cached"]], raw=["dynamic", 0]
+        ).positive
+        self.assertEqual(receipts.calls, [(JOB, "p", (("cached",),))])
+        self.assertEqual(result.raw.value, "LLM raw")
+        self.assertEqual(
+            (result.raw.provenance, result.raw.source), ("inferred", "graph_semantics")
+        )
+        self.assertEqual(result.effective.value, "cached")
+        self.assertEqual(result.effective.source, "host_cache_output")
+        # Fields the graph already knows are never replaced
+        self.assertEqual(result.prepend.value, " pre ")
+        self.assertEqual(result.prepend.source, "graph_literal")
+        self.assertEqual(result.append.value, " post ")
+
+    def test_cached_runtime_prepend_comes_from_the_matching_receipt(self):
+        receipts = self.Receipts([((" raw ", "LLM pre", " post "),)])
+        result = self.observe_with(
+            receipts, [object(), ["cached"]], prepend=["dynamic", 0]
+        ).positive
+        self.assertEqual(result.prepend.value, "LLM pre")
+        self.assertEqual(result.prepend.source, "graph_semantics")
+        self.assertEqual(result.raw.value, " raw ")
+        self.assertEqual(result.raw.source, "graph_literal")
+
+    def test_a_receipt_disagreeing_with_a_known_field_fills_nothing(self):
+        receipts = self.Receipts([(("LLM raw", "other pre", " post "),)])
+        result = self.observe_with(
+            receipts, [object(), ["cached"]], raw=["dynamic", 0]
+        ).positive
+        self.assertIsNone(result.raw.value)
+        self.assertEqual(result.prepend.value, " pre ")
+
+    def test_pair_runtime_negative_takes_the_negative_fields(self):
+        receipts = self.Receipts(
+            [((" raw ", " pre ", " post "), ("LLM negative", "", ""))]
+        )
+        result = self.observe_with(
+            receipts,
+            [object(), object(), ["cached positive"], ["LLM negative"]],
+            kind="PromptManagerPair",
+            negative=["dynamic", 0],
+        )
+        self.assertEqual(result.negative.raw.value, "LLM negative")
+        self.assertEqual(result.negative.raw.source, "graph_semantics")
+        self.assertEqual(result.positive.raw.value, " raw ")
+        self.assertEqual(result.positive.raw.source, "graph_literal")
+
+    def test_known_typed_fields_are_never_replaced_by_a_receipt(self):
+        receipts = self.Receipts([(("other", "other", "other"),)])
+        result = self.observe_with(receipts, [object(), ["cached"]]).positive
+        self.assertEqual(
+            (result.raw.value, result.prepend.value, result.append.value),
+            (" raw ", " pre ", " post "),
+        )
+        self.assertEqual(result.raw.source, "graph_literal")
+
+    def test_no_matching_receipt_keeps_the_cached_raw_unknown(self):
+        for fields in (None, [None], [], "bad", [(("a", "b"),)], [((1, "", ""),)]):
+            with self.subTest(fields=fields):
+                result = self.observe_with(
+                    self.Receipts(fields), [object(), ["cached"]], raw=["dynamic", 0]
+                ).positive
+                self.assertIsNone(result.raw.value)
+                self.assertEqual(result.effective.value, "cached")
+
+    def test_a_failing_receipt_lookup_keeps_the_capture(self):
+        class Broken:
+            def recorded_fields(self, *args):
+                raise RuntimeError("boom")
+
+        result = self.observe_with(
+            Broken(), [object(), ["cached"]], raw=["dynamic", 0]
+        ).positive
+        self.assertIsNone(result.raw.value)
+        self.assertEqual(result.effective.value, "cached")
+
+    def test_pair_slots_take_their_own_receipt_fields(self):
+        receipts = self.Receipts(
+            [(("LLM positive", " pre ", " post "), ("LLM negative", "", ""))]
+        )
+        result = self.observe_with(
+            receipts,
+            [object(), object(), ["cached positive"], ["cached negative"]],
+            kind="PromptManagerPair",
+            raw=["dynamic", 0],
+        )
+        self.assertEqual(
+            receipts.calls[0], (JOB, "p", (("cached positive",), ("cached negative",)))
+        )
+        self.assertEqual(result.positive.raw.value, "LLM positive")
+        self.assertEqual(result.positive.effective.value, "cached positive")
+        # The negative editor is typed: the graph's value stays
+        self.assertEqual(result.negative.raw.value, "negative")
+        self.assertEqual(result.negative.raw.source, "graph_literal")
+
+    def test_default_receipts_are_the_process_cached_runs(self):
+        from utils import cached_runs
+
+        receipts = self.Receipts([(("LLM raw", " pre ", " post "),)])
+        with patch.object(cached_runs, "CACHED_RUNS", receipts):
+            self.begin(raw=["dynamic", 0])
+            self.executor.caches.outputs.entries["p"] = Entry(
+                None, [object(), ["cached"]]
+            )
+            self.observe("execution_start")
+            self.observe("execution_cached", nodes=["p"])
+        self.assertEqual(self.seal().positive.raw.value, "LLM raw")

tests/test_execution_history_hooks.py

@@ -54,7 +54,7 @@ def status(kind="execution_success", job=JOB):
     return Status("success", True, [(kind, {"prompt_id": job, "timestamp": 20})])
 
 
-class HistoryHooksTests(unittest.TestCase):
+class HooksCase(unittest.TestCase):
     def setUp(self):
         self.directory = tempfile.TemporaryDirectory()
         self.addCleanup(self.directory.cleanup)
@@ -122,6 +122,8 @@ class HistoryHooksTests(unittest.TestCase):
     def drain(self, registration):
         drain_writer(registration.writer, clock=time.monotonic)
 
+
+class HistoryHooksTests(HooksCase):
     def test_package_history_initializes_when_monitor_and_usage_setup_fail(self):
         try:
             from tests.test_init_package import (
@@ -767,3 +769,150 @@ class HistoryHooksTests(unittest.TestCase):
         self.assertEqual(
             calls, ["execution_start", "execution_cached", "execution_cached"]
         )
+
+
+class CachedRuntimeHistoryTests(HooksCase):
+    """A cached runtime-linked prompt is a tracked run, like an executed one (#403).
+
+    The real history hook and writer, a real ``CachedRuns`` and the real host
+    adapter over fakes of ComfyUI's executor, cache and executing context.
+    """
+
+    TEXT = "a lighthouse the LLM wrote"
+    GRAPH = {
+        "str": {"class_type": "PrimitiveString", "inputs": {"value": "seed"}},
+        "llm": {"class_type": "PreviewAny", "inputs": {"source": ["str", 0]}},
+        "p": {"class_type": "PromptManager", "inputs": {"text": ["llm", 0]}},
+        "s": {"class_type": "KSampler", "inputs": {"positive": ["p", 0]}},
+    }
+
+    def setUp(self):
+        super().setUp()
+        from utils import cached_runs
+        from utils.cached_runs import CachedRuns
+
+        self.context = None
+        self.runs = CachedRuns(context=lambda: self.context or (None, None))
+        self.cache_entries = {}
+        entries = self.cache_entries
+        execution = types.ModuleType("execution")
+        caching = types.ModuleType("comfy_execution.caching")
+        package = types.ModuleType("comfy_execution")
+        utils_module = types.ModuleType("comfy_execution.utils")
+        entry_type = namedtuple("CacheEntry", "ui outputs")
+
+        class PromptExecutor:
+            def add_message(self, event, data, broadcast):
+                return None
+
+        class HierarchicalCache:
+            def get_local(self, node_id):
+                return entries.get(node_id)
+
+        execution.PromptExecutor, execution.CacheEntry = PromptExecutor, entry_type
+        caching.HierarchicalCache = HierarchicalCache
+        package.caching, package.utils = caching, utils_module
+        utils_module.get_executing_context = lambda: (
+            SimpleNamespace(prompt_id=self.context[0], node_id="p", list_index=0)
+            if self.context
+            else None
+        )
+        self.entry_type = entry_type
+        modules = patch.dict(
+            sys.modules,
+            {
+                "execution": execution,
+                "comfy_execution": package,
+                "comfy_execution.caching": caching,
+                "comfy_execution.utils": utils_module,
+            },
+        )
+        modules.start()
+        self.addCleanup(modules.stop)
+        swap = patch.object(cached_runs, "CACHED_RUNS", self.runs)
+        swap.start()
+        self.addCleanup(swap.stop)
+        self.executor = PromptExecutor()
+        self.executor.server = self.server
+        self.executor.caches = SimpleNamespace(outputs=HierarchicalCache())
+        self.prompt_id = self.db.save_prompt(text=self.TEXT)
+        self.item_id = 0
+
+    def job(self, cached):
+        """One job; the node executes unless ``cached`` (ComfyUI's cache serves it)."""
+        from uuid import uuid4
+
+        from utils.cached_runs import stored_facts
+        from utils.execution_context_capture import record_runtime_context
+
+        self.item_id += 1
+        job = str(uuid4())
+        result = ((0, job, self.GRAPH, {}, ["s"]), self.item_id)
+        self.queue.results.append(result)
+        self.queue.get()
+        self.runs.begin_attempt(result)
+        self.context = (job, 0)
+        self.executor.add_message("execution_start", {"prompt_id": job}, False)
+        if cached:
+            self.executor.add_message(
+                "execution_cached", {"prompt_id": job, "nodes": ["p"]}, False
+            )
+        else:
+            # What PromptManager.encode_prompt records for this text
+            record_runtime_context("PromptManager", self.TEXT, "", "", self.TEXT)
+            self.runs.record_item(
+                "p", (self.TEXT,), stored_facts(self.TEXT), ((self.TEXT, "", ""),)
+            )
+            self.cache_entries["p"] = self.entry_type(None, [object(), [self.TEXT]])
+        self.context = None
+        self.queue.task_done(self.item_id, {}, status(job=job))
+        self.runs.end_attempt(self.item_id)
+
+    def tracked_runs(self):
+        with self.db.recorded_history_snapshot():
+            return self.db.list_history_window(self.prompt_id)["recorded_run_count"]
+
+    def test_three_jobs_of_runtime_text_are_three_tracked_runs(self):
+        registration = self.register()
+        for cached in (False, True, True):
+            self.job(cached)
+        self.drain(registration)
+        self.assertEqual(self.tracked_runs(), 3)
+        rows = self.rows()
+        for row in rows:
+            contexts = self.db.get_execution_contexts(row["run_id"])["contexts"]
+            self.assertEqual(contexts[0]["positive"]["raw"], self.TEXT)
+            self.assertEqual(contexts[0]["positive_prompt_id"], self.prompt_id)
+        cached = self.db.get_execution_contexts(rows[1]["run_id"])["contexts"][0]
+        self.assertEqual(
+            cached["evidence"]["positive"]["raw"],
+            {"provenance": "inferred", "source": "graph_semantics", "reason": None},
+        )
+        self.assertEqual(
+            cached["evidence"]["positive"]["effective"]["source"], "host_cache_output"
+        )
+        # The stored history passes the whole-database check older builds run too
+        from database.execution_history_schema import history_schema_error
+
+        self.assertIsNone(history_schema_error(self.db.model.get_connection()))
+        # An image of a cached job can now be attributed through its run (#342)
+        from database.run_attribution import named_prompt
+
+        self.assertEqual(
+            [named_prompt(self.db, row["run_id"]) for row in rows],
+            [self.prompt_id] * 3,
+        )
+
+    def test_without_a_receipt_a_cached_job_stays_untracked(self):
+        registration = self.register()
+        self.job(False)
+        self.runs._receipts.clear()  # as after a restart with a provider entry
+        self.job(True)
+        self.drain(registration)
+        self.assertEqual(self.tracked_runs(), 1)
+        cached = self.db.get_execution_contexts(self.rows()[1]["run_id"])["contexts"]
+        self.assertIsNone(cached[0]["positive"]["raw"])
+        self.assertIsNone(cached[0]["positive_prompt_id"])
+        from database.run_attribution import named_prompt
+
+        self.assertIsNone(named_prompt(self.db, self.rows()[1]["run_id"]))

tests/test_pair_usage_tracking.py

@@ -99,7 +99,10 @@ class TestPairUsageTracking(unittest.TestCase):
             unique_id="1",
         )
         node.cached_runs.record_item.assert_called_once_with(
-            "1", ("cat", ""), Facts(usage_tracking.generate_prompt_hash("cat"), None)
+            "1",
+            ("cat", ""),
+            Facts(usage_tracking.generate_prompt_hash("cat"), None),
+            (("cat", "", ""), ("", "", "")),
         )
 
     def test_validation_rejection_never_counts_or_saves(self):

tests/test_prompt_usage.py

@@ -308,7 +308,11 @@ class TestNodeRecordsCachedRunReceipt(UsageTestCase):
             "typed text", prepend_text="pre", prompt=self.GRAPH, unique_id="1"
         )
         node.cached_runs.record_item.assert_called_once_with(
-            "1", ("pre typed text",), Facts(generate_prompt_hash("typed text"))
+            "1",
+            ("pre typed text",),
+            Facts(generate_prompt_hash("typed text")),
+            # #403: the texts the history capture got, for a cached job's history
+            (("typed text", "pre", ""),),
         )
         # Hook-visible text: the node counts nothing itself
         row = self.db.get_prompt_by_hash(generate_prompt_hash("typed text"))
@@ -319,7 +323,9 @@ class TestNodeRecordsCachedRunReceipt(UsageTestCase):
 
         node = self._node()
         node.process_text("  ", prompt=self.GRAPH, unique_id="1")
-        node.cached_runs.record_item.assert_called_once_with("1", ("  ",), Facts(None))
+        node.cached_runs.record_item.assert_called_once_with(
+            "1", ("  ",), Facts(None), (("  ", "", ""),)
+        )
 
 
 if __name__ == "__main__":

utils/cached_runs.py

@@ -15,9 +15,14 @@ role (positive, negative, which Pair outputs are used) and what the hook
 already saw are recomputed from that job's own graph, and both paths count
 through the same ``own_counts`` so they cannot drift.
 
+The receipt also keeps the raw, prepend and append text the node gave the
+execution history (#403), so the history capture of a cached job can record
+the raw text the graph cannot show (``recorded_fields``) and the job becomes a
+tracked run of its prompt, as it is a counted one.
+
 Fail closed: when the cached text cannot be matched to a receipt recorded in
-this process, nothing is counted (the behavior before #347). In memory only and
-bounded; starts no threads.
+this process, nothing is counted (the behavior before #347) and the history
+fields stay unknown. In memory only and bounded; starts no threads.
 """
 
 import hashlib
@@ -39,6 +44,7 @@ try:
     from .queue_observers import subscribe_dequeue, subscribe_task_done
     from .run_keys import prompt_node_key, upstream_nodes
     from .usage_tracking import find_pair_id_by_hashes, survive_switch
+    from .execution_contexts import MAX_ATTEMPT_BYTES, MAX_FIELD_BYTES, text_size
 except ImportError:
     from utils.hashing import generate_prompt_hash
     from utils.logging_config import get_logger
@@ -53,6 +59,7 @@ except ImportError:
     from utils.queue_observers import subscribe_dequeue, subscribe_task_done
     from utils.run_keys import prompt_node_key, upstream_nodes
     from utils.usage_tracking import find_pair_id_by_hashes, survive_switch
+    from utils.execution_contexts import MAX_ATTEMPT_BYTES, MAX_FIELD_BYTES, text_size
 
 logger = get_logger("prompt_manager.cached_runs")
 
@@ -61,9 +68,15 @@ MAX_KEYS = 256  # run keys remembered (the same bound as MAX_KEYED_RUNS)
 MAX_RECEIPTS = 4  # receipts per run key: a recurring upstream value (A, B, A)
 MAX_ITEMS = 64  # list items per receipt (the cache reader's bound)
 MAX_CACHED_NODES = 500  # the execution_cached guard the history capture uses
+MAX_FIELD_TEXT_BYTES = 8 * 1024 * 1024  # history fields all receipts keep
+MAX_RECEIPT_FIELD_BYTES = MAX_ATTEMPT_BYTES  # one receipt: what one run's history holds
 _EMPTY_HASH = generate_prompt_hash("")
 
 Count = Tuple[Any, ...]  # ("prompt", hash) or ("pair", positive, negative or None)
+# (raw, prepend, append) per STRING slot, as given to record_runtime_context
+Fields = Tuple[Tuple[str, str, str], ...]
+# Executions with equal outputs recorded different fields: none can be named
+_CONFLICT: Fields = ()
 
 
 class Facts(NamedTuple):
@@ -96,6 +109,8 @@ class Seen(NamedTuple):
 class _Item(NamedTuple):
     outputs: Tuple[str, ...]
     facts: Facts
+    fields: Optional[Fields] = None
+    field_bytes: int = 0
 
 
 class _Receipt:
@@ -108,6 +123,18 @@ class _Receipt:
     def outputs(self) -> Dict[int, Tuple[str, ...]]:
         return {index: item.outputs for index, item in self.items.items()}
 
+    def field_bytes(self) -> int:
+        return sum(item.field_bytes for item in self.items.values())
+
+    def strip_fields(self) -> int:
+        """Drop the history fields (counting is unchanged); the bytes freed."""
+        freed = 0
+        for index, item in list(self.items.items()):
+            if item.field_bytes:
+                freed += item.field_bytes
+                self.items[index] = item._replace(fields=None, field_bytes=0)
+        return freed
+
 
 class _Attempt(NamedTuple):
     item_id: int
@@ -117,6 +144,22 @@ class _Attempt(NamedTuple):
     replayed: set
 
 
+def _field_size(fields: Any, slots: int) -> Optional[int]:
+    """UTF-8 size of well-formed fields for ``slots`` outputs, else None."""
+    if type(fields) is not tuple or len(fields) != slots:
+        return None
+    size = 0
+    for slot in fields:
+        if type(slot) is not tuple or len(slot) != 3:
+            return None
+        for value in slot:
+            try:
+                size += text_size(value, MAX_FIELD_BYTES)  # the capture's own check
+            except ValueError:
+                return None
+    return size
+
+
 def text_digest(value: str) -> str:
     """Digest of one returned STRING value, exactly as returned."""
     return hashlib.sha256(value.encode("utf-8", "surrogatepass")).hexdigest()
@@ -245,11 +288,14 @@ class CachedRuns:
         max_keys: int = MAX_KEYS,
         max_receipts: int = MAX_RECEIPTS,
         max_items: int = MAX_ITEMS,
+        max_field_bytes: int = MAX_FIELD_TEXT_BYTES,
         context: Callable[[], Tuple[Optional[str], Optional[int]]] = _host_context,
     ) -> None:
         self._max_keys = max_keys
         self._max_receipts = max_receipts
         self._max_items = max_items
+        self._max_field_bytes = max_field_bytes
+        self._field_bytes = 0  # UTF-8 bytes of the fields all receipts keep
         self._context = context
         self._receipts: "OrderedDict[str, List[_Receipt]]" = OrderedDict()
         self._attempt: Optional[_Attempt] = None
@@ -284,7 +330,10 @@ class CachedRuns:
                 self._attempt = None
             for run_key, receipts in self._receipts.items():
                 if receipts and receipts[0].item_id == item_id:
-                    self._receipts[run_key] = _deduplicated(receipts)
+                    before = _bytes(receipts)
+                    deduplicated = _deduplicated(receipts)
+                    self._field_bytes += _bytes(deduplicated) - before
+                    self._receipts[run_key] = deduplicated
 
     def _current(self) -> Tuple[Optional[_Attempt], Optional[int]]:
         """The attempt the executing node belongs to, and its list index."""
@@ -305,8 +354,19 @@ class CachedRuns:
         with self._lock:
             return str(node_id) in attempt.replayed
 
-    def record_item(self, node_id: Any, outputs: Tuple[Any, ...], facts: Facts) -> None:
-        """Remember what this execution returned and was entitled to count."""
+    def record_item(
+        self,
+        node_id: Any,
+        outputs: Tuple[Any, ...],
+        facts: Facts,
+        fields: Optional[Fields] = None,
+    ) -> None:
+        """Remember what this execution returned and was entitled to count.
+
+        ``fields`` are the texts the node gave the execution history, one
+        ``(raw, prepend, append)`` per output; kept only when well formed,
+        each within the history's field limit and the store within its budget.
+        """
         try:
             attempt, index = self._current()
             node = attempt.nodes.get(str(node_id)) if attempt else None
@@ -317,14 +377,20 @@ class CachedRuns:
                 return
             if any(not isinstance(value, str) for value in outputs):
                 return
-            item = _Item(tuple(text_digest(value) for value in outputs), facts)
+            size = _field_size(fields, len(outputs)) if fields is not None else None
+            item = _Item(
+                tuple(text_digest(value) for value in outputs),
+                facts,
+                fields if size is not None else None,
+                size or 0,
+            )
             self._store(node[1], attempt.item_id, index, item)
         except Exception as error:
             logger.debug(f"Could not record a cached-run receipt: {error}")
 
     def _store(self, run_key: str, item_id: int, index: int, item: _Item) -> None:
         with self._lock:
-            receipts = self._receipts.pop(run_key, [])
+            receipts = self._take(run_key)
             if receipts and receipts[0].item_id == item_id:
                 receipt = receipts[0]
             else:
@@ -334,9 +400,44 @@ class CachedRuns:
                 receipt = _Receipt(item_id)
                 receipts.insert(0, receipt)
             receipt.items[index] = item
-            self._receipts[run_key] = receipts[: self._max_receipts]
+            receipts = receipts[: self._max_receipts]
+            if item.field_bytes and (
+                receipt.field_bytes() > MAX_RECEIPT_FIELD_BYTES
+                or not self._make_room(_bytes(receipts))
+            ):
+                # Counting is as before; this job's history fields stay unknown
+                receipt.items[index] = item._replace(fields=None, field_bytes=0)
+            self._put(run_key, receipts)
             while len(self._receipts) > self._max_keys:
-                self._receipts.popitem(last=False)
+                _, dropped = self._receipts.popitem(last=False)
+                self._field_bytes -= _bytes(dropped)
+
+    def _make_room(self, needed: int) -> bool:
+        """Free the fields of the least recently used receipts (under the lock).
+
+        Old receipts of workflows never run again must not hold the budget:
+        their counting stays, only their history fields go. False when
+        ``needed`` does not fit even then.
+        """
+        if needed > self._max_field_bytes:
+            return False
+        for receipts in self._receipts.values():  # least recently used first
+            if self._field_bytes + needed <= self._max_field_bytes:
+                break
+            for receipt in receipts:
+                self._field_bytes -= receipt.strip_fields()
+        return self._field_bytes + needed <= self._max_field_bytes
+
+    def _take(self, run_key: str) -> List[_Receipt]:
+        """Remove a key's receipts and their bytes from the store (under the lock)."""
+        receipts = self._receipts.pop(run_key, [])
+        self._field_bytes -= _bytes(receipts)
+        return list(receipts)
+
+    def _put(self, run_key: str, receipts: List[_Receipt]) -> None:
+        """Store a key's receipts as the most recently used (under the lock)."""
+        self._receipts[run_key] = receipts
+        self._field_bytes += _bytes(receipts)
 
     # -- cached nodes -------------------------------------------------------
 
@@ -393,25 +494,11 @@ class CachedRuns:
 
     def _match(self, run_key: str, values: Any) -> Optional[List[_Item]]:
         """The items of the one receipt whose outputs equal ``values``, in order."""
-        if not values:
-            return None
-        length = len(values[0])
-        if length == 0 or length > self._max_items:
-            return None
-        if any(len(slot) != length for slot in values):
-            return None
-        wanted = {
-            index: tuple(text_digest(slot[index]) for slot in values)
-            for index in range(length)
-        }
         with self._lock:
-            receipts = self._receipts.get(run_key)
-            if not receipts:
-                return None
-            matches = [receipt for receipt in receipts if receipt.outputs() == wanted]
+            matches = self._matching(run_key, values)
             if not matches:
                 return None
-            self._receipts.move_to_end(run_key)
+            length = len(values[0])
             facts = {
                 tuple(receipt.items[index].facts for index in range(length))
                 for receipt in matches
@@ -420,13 +507,99 @@ class CachedRuns:
                 return None
             return [matches[0].items[index] for index in range(length)]
 
+    def _matching(self, run_key: str, values: Any) -> List[_Receipt]:
+        """Receipts of ``run_key`` whose outputs equal ``values`` (under the lock)."""
+        if type(values) not in (list, tuple) or not values:
+            return []
+        if any(type(slot) not in (list, tuple) for slot in values):
+            return []
+        length = len(values[0])
+        if length == 0 or length > self._max_items:
+            return []
+        if any(len(slot) != length for slot in values):
+            return []
+        if any(type(value) is not str for slot in values for value in slot):
+            return []
+        wanted = {
+            index: tuple(text_digest(slot[index]) for slot in values)
+            for index in range(length)
+        }
+        receipts = self._receipts.get(run_key)
+        if not receipts:
+            return []
+        matches = [receipt for receipt in receipts if receipt.outputs() == wanted]
+        if matches:
+            self._receipts.move_to_end(run_key)
+        return matches
+
+    # -- the history capture (#403) -------------------------------------------
+
+    def recorded_fields(
+        self, prompt_id: Any, node_id: Any, values: Any
+    ) -> Optional[List[Fields]]:
+        """The history fields of the execution a cached node reuses, per list item.
+
+        ``values`` are the node's cached STRING outputs (``cached_string_outputs``).
+        Only for a node of the job currently executing, and only when every
+        receipt matching ``values`` kept the same fields; None otherwise. Never
+        counts and never marks the node replayed.
+        """
+        try:
+            with self._lock:
+                attempt = self._attempt
+                if (
+                    attempt is None
+                    or type(prompt_id) is not str
+                    or prompt_id != attempt.prompt_id
+                    or type(node_id) is not str
+                    or node_id not in attempt.nodes
+                ):
+                    return None
+                matches = self._matching(attempt.nodes[node_id][1], values)
+                if not matches:
+                    return None
+                found = []
+                for index in range(len(values[0])):
+                    fields = {receipt.items[index].fields for receipt in matches}
+                    if len(fields) != 1 or None in fields or _CONFLICT in fields:
+                        return None
+                    found.append(fields.pop())
+                return found
+        except Exception as error:
+            logger.debug(f"Could not read a cached run's history fields: {error}")
+            return None
+
 
 def _deduplicated(receipts: List[_Receipt]) -> List[_Receipt]:
-    """The newest receipt replaces older ones with the same outputs."""
+    """The newest receipt replaces older ones with the same outputs.
+
+    Counting needs only the outputs. When a replaced receipt recorded other
+    history fields for equal outputs (a raw text the outputs do not show),
+    the newest item's fields become a conflict, so no cached job of it is
+    given either text (#403). A conflict stays until the run key is dropped.
+    """
     if not receipts:
         return receipts
-    outputs = receipts[0].outputs()
-    return [receipts[0]] + [r for r in receipts[1:] if r.outputs() != outputs]
+    newest, outputs = receipts[0], receipts[0].outputs()
+    kept = [newest]
+    for receipt in receipts[1:]:
+        if receipt.outputs() != outputs:
+            kept.append(receipt)
+            continue
+        for index, item in list(newest.items.items()):
+            older = receipt.items[index].fields
+            if older == _CONFLICT or (
+                older is not None and item.fields is not None and older != item.fields
+            ):
+                newest.items[index] = item._replace(fields=_CONFLICT, field_bytes=0)
+    return kept
+
+
+def _bytes(receipts: List[_Receipt]) -> int:
+    """UTF-8 bytes of the history fields ``receipts`` keep."""
+    return sum(
+        item.field_bytes for receipt in receipts for item in receipt.items.values()
+    )
 
 
 def _job_nodes(graph: Dict[str, Any], outputs: Any) -> Dict[str, Tuple[str, str]]:

utils/execution_context_host.py

@@ -151,10 +151,77 @@ def cached_string_outputs(
         return None
 
 
+def _receipt_fields(
+    receipts: object, executor: object, job_id: str, source: object
+) -> list | None:
+    """Per list index, the ``(raw, prepend, append)`` of the execution reused (#403).
+
+    From the #347 receipt whose digests equal every STRING slot of the cache
+    entry; None when there is none or the answer is malformed.
+    """
+    try:
+        if receipts is None:
+            try:
+                from .cached_runs import CACHED_RUNS as receipts
+            except ImportError:
+                from utils.cached_runs import CACHED_RUNS as receipts
+        values = cached_string_outputs(executor, source.node_id, source.node_type)
+        if values is None:
+            return None
+        found = receipts.recorded_fields(job_id, source.node_id, values)
+        if type(found) is not list or len(found) != len(values[0]):
+            return None
+        position = (
+            (1 if source.output_slot in (1, 3) else 0)
+            if source.node_type == "PromptManagerPair"
+            else 0
+        )
+        fields = []
+        for slots in found:
+            slot = slots[position] if type(slots) is tuple else None
+            if not (
+                type(slot) is tuple
+                and len(slot) == 3
+                and all(type(value) is str for value in slot)
+            ):
+                return None
+            fields.append(slot)
+        return fields
+    except Exception:
+        return None
+
+
+def _agrees(inferred: PromptSnapshot, fields: tuple) -> bool:
+    """Every field the graph knows equals the receipt's; else fill nothing."""
+    known = (inferred.raw, inferred.prepend, inferred.append)
+    return all(
+        field.value is None or field.value == value
+        for field, value in zip(known, fields)
+    )
+
+
+def _filled(inferred: TextEvidence, value: str) -> TextEvidence:
+    """A field the graph cannot show, taken from the reused execution (#403)."""
+    if inferred.value is not None:
+        return inferred
+    return TextEvidence(value, "inferred", "graph_semantics", None)
+
+
 def observe_cached_context(
-    executor: object, message_type: object, message_data: object
+    executor: object,
+    message_type: object,
+    message_data: object,
+    *,
+    receipts: object = None,
 ) -> None:
-    """Copy bounded local strings without cache/provider execution or retention."""
+    """Copy bounded local strings without cache/provider execution or retention.
+
+    Raw, prepend and append come from graph inference; a field it leaves
+    unknown (runtime-only text) is taken from the receipt of the execution the
+    cache entry is proven to come from (``CachedRuns.recorded_fields``), so the
+    cached job is a tracked run of its prompt (#403). ``receipts`` defaults to
+    the process ``CACHED_RUNS``.
+    """
     try:
         binding = active_capture()
         if binding is None or type(message_type) is not str:
@@ -198,18 +265,29 @@ def observe_cached_context(
             inferred = manager.source_snapshot(cap, source)
             if inferred is None:
                 continue
+            fields = None
+            if any(
+                field.value is None
+                for field in (inferred.raw, inferred.prepend, inferred.append)
+            ):
+                fields = _receipt_fields(receipts, executor, cap.job_id, source)
             for index, value in enumerate(values):
                 if type(value) is not str:
                     continue
+                raw, prepend, append = inferred.raw, inferred.prepend, inferred.append
+                if fields is not None and _agrees(inferred, fields[index]):
+                    raw = _filled(raw, fields[index][0])
+                    prepend = _filled(prepend, fields[index][1])
+                    append = _filled(append, fields[index][2])
                 manager.observe(
                     cap,
                     source,
                     index,
                     PromptSnapshot(
-                        inferred.raw,
+                        raw,
                         TextEvidence(value, "recorded", "host_cache_output", None),
-                        inferred.prepend,
-                        inferred.append,
+                        prepend,
+                        append,
                     ),
                 )
     except Exception: