From fa16578a9732e59e1be2bdc8f55092beb1014d6d Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Fri, 18 Sep 2026 00:54:25 +0000
Subject: [PATCH 1/2] Always overwrite on Generate/Summarize/Describe; fix
 collateral provenance lock

Pressing Generate/Summarize/Describe is an explicit request for a fresh
answer, so apply_ai_metadata no longer skips a field marked "human" in
field_provenance -- it always overwrites, on the single-asset buttons,
"Generate all", and bulk actions alike (they all route through the same
summarize.run/describe.run). field_provenance is still stamped for
bookkeeping, just no longer read as a gate.

Also fixes the actual reason fields were getting silently locked in the
first place: AssetDetail's save button sent name/description/summary as a
fixed trio whenever any one changed, so editing just the name also
re-stamped the other two as human-written -- even while empty. It now
sends only the fields that actually changed.

summarize and describe also now feed the asset's own tags into the prompt
as additional context, alongside the filename and existing material.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FdoHBNLsz2cXKBVzsqKPoV
---
 backend/app/enrichment/describe.py           | 12 +++--
 backend/app/enrichment/summarize.py          | 24 +++++----
 backend/app/services/assets.py               | 36 +++++--------
 backend/tests/test_assets_api.py             |  3 +-
 backend/tests/test_describe.py               | 25 +++++++--
 backend/tests/test_summarize.py              | 49 ++++++++++--------
 backend/tests/test_usage.py                  | 16 ------
 docs/m6-ai-enrichment.md                     | 54 +++++++++++++-------
 docs/plan-of-attack.md                       |  7 ++-
 frontend/src/components/AssetDetail.tsx      | 17 +++---
 frontend/src/components/EnrichmentButton.tsx |  3 +-
 11 files changed, 135 insertions(+), 111 deletions(-)

diff --git a/backend/app/enrichment/describe.py b/backend/app/enrichment/describe.py
index 8fcef9f..6b1d21d 100644
--- a/backend/app/enrichment/describe.py
+++ b/backend/app/enrichment/describe.py
@@ -27,6 +27,7 @@
 from app.providers import build_provider
 from app.providers.base import ProviderError, ProviderUnavailable
 from app.services import assets as asset_service
+from app.services import tags as tag_service
 from app.usage import events as usage_events
 
 logger = logging.getLogger(__name__)
@@ -55,7 +56,7 @@
 )
 
 
-def _prompt(asset: Asset, material: source.SourceMaterial) -> str:
+def _prompt(asset: Asset, material: source.SourceMaterial, tag_names: list[str]) -> str:
     parts = [f"Filename: {asset.original_name or asset.name}"]
     if asset.summary:
         # Context, not something to restate — saying so is what stops the model
@@ -64,6 +65,8 @@ def _prompt(asset: Asset, material: source.SourceMaterial) -> str:
             f"A summary already exists; do not repeat it, describe what is in the "
             f"item instead: {asset.summary}"
         )
+    if tag_names:
+        parts.append("Tags already applied to this item: " + ", ".join(tag_names))
 
     if material.kind == source.FROM_TRANSCRIPT:
         header = "Transcript of the recording"
@@ -109,8 +112,9 @@ def run(session: Session, asset: Asset, progress: Progress) -> str:
     )
 
     progress("Describing", 30, "")
+    tag_names = [t.name for t in tag_service.tags_for(session, asset.id)]
     completion = provider.complete(
-        _prompt(asset, material),
+        _prompt(asset, material, tag_names),
         system=SYSTEM_PROMPT,
         images=material.images,
     )
@@ -129,9 +133,7 @@ def run(session: Session, asset: Asset, progress: Progress) -> str:
     # land a description afterwards.
     progress("Saving", 90, "")
 
-    written = asset_service.apply_ai_metadata(session, asset, {"description": text})
-    if not written:
-        return "Left alone — you wrote this description yourself"
+    asset_service.apply_ai_metadata(session, asset, {"description": text})
 
     saw = " and a frame" if material.images and material.kind == source.FROM_TRANSCRIPT else ""
     logger.info(
diff --git a/backend/app/enrichment/summarize.py b/backend/app/enrichment/summarize.py
index 3cf099b..11faa90 100644
--- a/backend/app/enrichment/summarize.py
+++ b/backend/app/enrichment/summarize.py
@@ -1,8 +1,10 @@
 """The summarize job: turn what an asset contains into a paragraph about it.
 
 The first job in M6 to produce something a person reads, and the first AI write path in
-the app — so it is also the first caller of `apply_ai_metadata`, which is what keeps a
-re-run from replacing a summary somebody wrote by hand (FR 8.1.3).
+the app — so it is also the first caller of `apply_ai_metadata`. Pressing the button is
+an explicit request for a fresh summary, so a re-run always replaces whatever is there,
+including a summary somebody typed by hand — `apply_ai_metadata` still records that the
+result came from AI, but nothing stops it from landing.
 
 What it reads is not this module's decision; `enrichment/source.py` makes it once for
 every job that will need it. For a transcribed video that is the transcript.
@@ -20,6 +22,7 @@
 from app.providers import build_provider
 from app.providers.base import ProviderError, ProviderUnavailable
 from app.services import assets as asset_service
+from app.services import tags as tag_service
 from app.usage import events as usage_events
 
 logger = logging.getLogger(__name__)
@@ -49,11 +52,13 @@
 )
 
 
-def _prompt(asset: Asset, material: source.SourceMaterial) -> str:
+def _prompt(asset: Asset, material: source.SourceMaterial, tag_names: list[str]) -> str:
     parts = [f"Filename: {asset.original_name or asset.name}"]
     if asset.description:
         # A description the user wrote is context, not something to restate.
         parts.append(f"The owner's own note about it: {asset.description}")
+    if tag_names:
+        parts.append("Tags already applied to this item: " + ", ".join(tag_names))
 
     if material.kind == source.FROM_TRANSCRIPT:
         header = "Transcript of the recording"
@@ -96,15 +101,16 @@ def run(session: Session, asset: Asset, progress: Progress) -> str:
     )
     progress("Summarising", 30, detail)
 
+    tag_names = [t.name for t in tag_service.tags_for(session, asset.id)]
     completion = provider.complete(
-        _prompt(asset, material),
+        _prompt(asset, material, tag_names),
         system=SYSTEM_PROMPT,
         images=material.images,
     )
 
     # Recorded before this job decides what to do with the answer, so usage does not
-    # depend on the outcome: a reply that gets trimmed, or that provenance then refuses
-    # to write, cost the same tokens as one that lands.
+    # depend on the outcome: a reply that gets trimmed still cost the same tokens as one
+    # that does not.
     #
     # Not a complete guarantee, and the gap is worth knowing: a reply the *provider's*
     # parser rejects — an empty completion, unusable JSON — raises before there is a
@@ -123,11 +129,7 @@ def run(session: Session, asset: Asset, progress: Progress) -> str:
     # not still land a summary afterwards.
     progress("Saving", 90, "")
 
-    written = asset_service.apply_ai_metadata(session, asset, {"summary": text})
-    if not written:
-        # Not an error. The user wrote their own summary, which outranks this one — but
-        # saying so beats a job that reports success and changed nothing.
-        return "Left alone — you wrote this summary yourself"
+    asset_service.apply_ai_metadata(session, asset, {"summary": text})
 
     logger.info(
         "Summarised asset %s from %s (%d in / %d out tokens)",
diff --git a/backend/app/services/assets.py b/backend/app/services/assets.py
index a4b5f65..b9bd021 100644
--- a/backend/app/services/assets.py
+++ b/backend/app/services/assets.py
@@ -265,10 +265,11 @@ def delete_asset(session: Session, storage: LocalStorage, asset: Asset) -> None:
 def apply_metadata(session: Session, asset: Asset, changes: dict) -> Asset:
     """Apply a user's manual edits, recording that a human made them.
 
-    The provenance write is what FR 8.1.3 rests on: a later AI enrichment run reads it
-    to know which fields a person has touched, and must not overwrite those without
-    asking. Recording it here — at the only place manual edits happen — is what keeps
-    that guarantee from depending on every future caller remembering it.
+    Stamping "human" here still matters for `name`: a suggested title is only ever
+    applied through this path (`services/suggestions.py::accept`), and there is no
+    direct AI write path that could replace it afterwards. For `description` and
+    `summary` the stamp is now informational only — `apply_ai_metadata` no longer reads
+    it before overwriting either field.
     """
     if not changes:
         return asset
@@ -293,19 +294,15 @@ def apply_metadata(session: Session, asset: Asset, changes: dict) -> Asset:
 
 
 def apply_ai_metadata(session: Session, asset: Asset, changes: dict) -> list[str]:
-    """Write fields an enrichment job produced, without overwriting a person's work.
+    """Write fields an enrichment job produced.
 
-    The counterpart to `apply_metadata`, and the first thing to actually *read*
-    `field_provenance` — FR 8.1.3 requires that a later AI run never silently replaces
-    something someone typed, and until now the column was written and consulted by
-    nothing.
+    The counterpart to `apply_metadata`. Pressing a "Generate"/"Summarize"/"Describe"
+    control is a deliberate, explicit request for a fresh answer, so it always lands —
+    including over a value a person typed by hand. `field_provenance` is still stamped
+    "ai" afterwards, so the record of who wrote a field stays accurate even though
+    nothing here gates on it any more.
 
-    A field with no provenance entry has never been touched by a person, so it is free
-    to write: absence is the signal, and no third provenance value is needed for it. A
-    field marked "human" is skipped and named in the return value, so the caller can say
-    what it left alone rather than reporting a clean run that quietly did less.
-
-    Returns the fields actually written.
+    Returns the fields written (always every key in `changes`, once any are given).
     """
     if not changes:
         return []
@@ -315,16 +312,9 @@ def apply_ai_metadata(session: Session, asset: Asset, changes: dict) -> list[str
     except ValueError:
         provenance = {}
 
-    written: list[str] = []
     for field_name, value in changes.items():
-        if provenance.get(field_name) == "human":
-            continue
         setattr(asset, field_name, value)
         provenance[field_name] = "ai"
-        written.append(field_name)
-
-    if not written:
-        return []
 
     asset.field_provenance = json.dumps(provenance, sort_keys=True)
     asset.metadata_modified_date = utcnow()
@@ -333,7 +323,7 @@ def apply_ai_metadata(session: Session, asset: Asset, changes: dict) -> list[str
     session.commit()
     session.refresh(asset)
     _reindex(session, asset)
-    return written
+    return list(changes.keys())
 
 
 # ─── serialisation ───────────────────────────────────────────────────────────
diff --git a/backend/tests/test_assets_api.py b/backend/tests/test_assets_api.py
index 53c549b..8894d64 100644
--- a/backend/tests/test_assets_api.py
+++ b/backend/tests/test_assets_api.py
@@ -206,7 +206,8 @@ def test_patch_updates_metadata_and_records_human_provenance(library, session):
     assert response.status_code == 200
     assert response.json()["data"]["name"] == "Ocean at sunset"
 
-    # FR 8.1.3: a later AI run reads this to know what a person wrote.
+    # FR 8.1.3 bookkeeping: still recorded, though only `name` has anything left that
+    # reads it (a suggested title never overwrites one a person already chose).
     asset = session.get(Asset, created["id"])
     import json
 
diff --git a/backend/tests/test_describe.py b/backend/tests/test_describe.py
index 6e37b11..c64e886 100644
--- a/backend/tests/test_describe.py
+++ b/backend/tests/test_describe.py
@@ -17,8 +17,9 @@
 from app.providers import _upstream
 from app.providers.base import ProviderError
 from app.services import assets as asset_service
+from app.services import tags as tag_service
 
-from tests.test_summarize import _upload_image, add_transcript, configure_provider
+from tests.test_summarize import TEST_USER, _upload_image, add_transcript, configure_provider
 
 DESCRIPTION = "James Giordano speaking to camera in a panelled studio, discussing DARPA."
 
@@ -115,6 +116,18 @@ def test_an_existing_summary_is_context_not_material_to_repeat(library, session,
     assert "do not repeat it" in prompt
 
 
+def test_the_assets_own_tags_are_given_as_context(library, session, upstream):
+    created = _upload_image(library)
+    asset = session.get(Asset, created["id"])
+    tag = tag_service.get_or_create(session, TEST_USER, "DARPA")
+    tag_service.attach(session, asset.id, tag.id)
+    configure_provider(session)
+
+    describe.run(session, asset, lambda *a, **k: None)
+
+    assert "DARPA" in sent_prompt(upstream)
+
+
 # ─── the ordinary paths ──────────────────────────────────────────────────────
 
 
@@ -139,17 +152,19 @@ def test_the_description_lands_on_the_asset(library, session, upstream):
     assert json.loads(asset.field_provenance)["description"] == "ai"
 
 
-def test_it_refuses_to_overwrite_a_description_someone_wrote(library, session, upstream):
+def test_a_description_someone_wrote_is_replaced_when_asked(library, session, upstream):
+    """Pressing Describe is an explicit request for a fresh answer, so it always lands
+    — even over a description a person typed themselves."""
     created = _upload_image(library)
     asset = session.get(Asset, created["id"])
     asset_service.apply_metadata(session, asset, {"description": "Mine, thanks."})
     configure_provider(session)
 
-    detail = describe.run(session, asset, lambda *a, **k: None)
+    describe.run(session, asset, lambda *a, **k: None)
 
     session.refresh(asset)
-    assert asset.description == "Mine, thanks."
-    assert "you wrote this" in detail.lower()
+    assert asset.description == DESCRIPTION
+    assert json.loads(asset.field_provenance)["description"] == "ai"
 
 
 def test_an_over_long_reply_is_trimmed(library, session, upstream):
diff --git a/backend/tests/test_summarize.py b/backend/tests/test_summarize.py
index 6f9a806..d37a002 100644
--- a/backend/tests/test_summarize.py
+++ b/backend/tests/test_summarize.py
@@ -2,7 +2,8 @@
 
 Two behaviours matter more than the rest and have most of the cases below: what the job
 chooses to read (a transcribed video is summarised from its transcript, never from one
-frame of it), and what it refuses to overwrite (a summary somebody typed).
+frame of it), and that pressing the button always lands a fresh answer — even over a
+summary somebody typed by hand.
 """
 
 import json
@@ -23,6 +24,7 @@
 from app.providers import _upstream
 from app.providers.base import ProviderError
 from app.services import assets as asset_service
+from app.services import tags as tag_service
 
 FIXTURES = Path(__file__).parent / "fixtures"
 TEST_USER = "user-under-test"
@@ -225,6 +227,18 @@ def test_the_owners_own_note_is_given_as_context(library, session, upstream):
     assert "Shot on the roof in Lisbon." in sent_prompt(upstream)
 
 
+def test_the_assets_own_tags_are_given_as_context(library, session, upstream):
+    created = _upload_image(library)
+    asset = session.get(Asset, created["id"])
+    tag = tag_service.get_or_create(session, TEST_USER, "DARPA")
+    tag_service.attach(session, asset.id, tag.id)
+    configure_provider(session)
+
+    summarize.run(session, asset, lambda *a, **k: None)
+
+    assert "DARPA" in sent_prompt(upstream)
+
+
 # ─── what it writes ──────────────────────────────────────────────────────────
 
 
@@ -240,19 +254,19 @@ def test_the_summary_lands_on_the_asset(library, session, upstream):
     assert json.loads(asset.field_provenance)["summary"] == "ai"
 
 
-def test_it_refuses_to_overwrite_a_summary_someone_wrote(library, session, upstream):
-    """FR 8.1.3. `field_provenance` was written from M1 and read by nothing until now."""
+def test_a_summary_someone_wrote_is_replaced_when_asked(library, session, upstream):
+    """Pressing Summarize is an explicit request for a fresh answer, so it always lands
+    — even over a summary a person typed themselves."""
     created = _upload_image(library)
     asset = session.get(Asset, created["id"])
     asset_service.apply_metadata(session, asset, {"summary": "Mine, thanks."})
     configure_provider(session)
 
-    detail = summarize.run(session, asset, lambda *a, **k: None)
+    summarize.run(session, asset, lambda *a, **k: None)
 
     session.refresh(asset)
-    assert asset.summary == "Mine, thanks."
-    # Reported rather than silently doing nothing and claiming success.
-    assert "you wrote this" in detail.lower()
+    assert asset.summary == SUMMARY
+    assert json.loads(asset.field_provenance)["summary"] == "ai"
 
 
 def test_a_previous_ai_summary_is_replaced(library, session, upstream):
@@ -302,8 +316,9 @@ def test_ai_metadata_writes_an_untouched_field(library, session):
     assert json.loads(asset.field_provenance) == {"summary": "ai"}
 
 
-def test_ai_metadata_skips_only_the_human_field(library, session):
-    """A mixed write should land the parts it is allowed to, not abort wholesale."""
+def test_ai_metadata_overwrites_a_field_marked_human(library, session):
+    """A field someone edited by hand is still fair game for the next AI write —
+    pressing Generate again is what asked for this."""
     created = _upload_image(library)
     asset = session.get(Asset, created["id"])
     asset_service.apply_metadata(session, asset, {"description": "Mine."})
@@ -312,21 +327,11 @@ def test_ai_metadata_skips_only_the_human_field(library, session):
         session, asset, {"description": "theirs", "summary": "auto"}
     )
 
-    assert written == ["summary"]
+    assert written == ["description", "summary"]
     session.refresh(asset)
-    assert asset.description == "Mine."
+    assert asset.description == "theirs"
     assert asset.summary == "auto"
-
-
-def test_ai_metadata_with_nothing_allowed_writes_nothing(library, session):
-    created = _upload_image(library)
-    asset = session.get(Asset, created["id"])
-    asset_service.apply_metadata(session, asset, {"summary": "Mine."})
-    before = asset.metadata_modified_date
-
-    assert asset_service.apply_ai_metadata(session, asset, {"summary": "auto"}) == []
-    session.refresh(asset)
-    assert asset.metadata_modified_date == before
+    assert json.loads(asset.field_provenance)["description"] == "ai"
 
 
 # ─── the endpoint ────────────────────────────────────────────────────────────
diff --git a/backend/tests/test_usage.py b/backend/tests/test_usage.py
index cfb59b5..1f7818e 100644
--- a/backend/tests/test_usage.py
+++ b/backend/tests/test_usage.py
@@ -142,22 +142,6 @@ def test_a_provider_reporting_no_tokens_records_nothing(library, session, upstre
     assert session.exec(select(UsageEvent)).all() == []
 
 
-def test_usage_does_not_depend_on_what_happens_to_the_answer(library, session, upstream):
-    """Provenance refuses the write, but the tokens were still spent. A total that only
-    counted runs whose output landed would understate what the library cost."""
-    from app.services import assets as asset_service
-
-    created = _upload_image(library)
-    asset = session.get(Asset, created["id"])
-    asset_service.apply_metadata(session, asset, {"summary": "Mine, thanks."})
-    configure_provider(session)
-
-    detail = summarize.run(session, asset, lambda *a, **k: None)
-
-    assert "you wrote this" in detail.lower()
-    assert len(session.exec(select(UsageEvent)).all()) == 1
-
-
 def test_a_reply_the_provider_parser_rejects_is_not_recorded(library, session, upstream):
     """The one gap, asserted rather than left to be discovered. An empty completion
     raises inside the client, before this job holds a Completion to record — so those
diff --git a/docs/m6-ai-enrichment.md b/docs/m6-ai-enrichment.md
index 2bfc0db..6265d27 100644
--- a/docs/m6-ai-enrichment.md
+++ b/docs/m6-ai-enrichment.md
@@ -210,18 +210,37 @@ Two decisions that go with it:
   user sees in their library without asking is a different act from filling a blank. It
   goes through the same accept/reject path as `autotag` (FR 9.1.4).
 
-### What protects a human edit
+### What `field_provenance` records
 
-`field_provenance` is written in exactly one place — `services/assets.py::apply_metadata`,
-which stamps `"human"` — and read nowhere. A freshly ingested asset therefore has `{}`.
-
-That absence is already the signal step 7 needs, so **no new provenance value is
-required**: a field with no entry has never been touched by a person and an AI run may
-write it; a field marked `"human"` may not be overwritten without asking. An AI write
-stamps `"ai"`, which a later AI run is free to replace.
+**Amended post-M6, once real use showed the original rule did not hold up. Read this in
+place of the "What protects a human edit" section it replaces; the rest of this document
+still reflects the design as landed.**
 
-`name` is the exception, and not because of provenance: it is never blank, so the
-suggestion rule above applies to it whether or not a person has edited it.
+`field_provenance` is written in exactly one place — `services/assets.py::apply_metadata`,
+which stamps `"human"` — and, as of M6, was read by `apply_ai_metadata` to *refuse* to
+overwrite a `"human"` field. That turned out to be the wrong default: pressing
+Summarize/Describe/"Generate all" is already a deliberate, explicit request for a fresh
+answer, and a silent no-op in response to it read as broken rather than protective —
+doubly so because the field it refused to touch was often one the person had never
+actually typed into (see below).
+
+So `apply_ai_metadata` no longer checks provenance before writing. Every enrichment write
+lands, unconditionally, including over a value a person wrote by hand. The column is
+still stamped — `"human"` from a manual edit, `"ai"` from a generated one — purely as a
+record of who last wrote a field, for the "surfacing provenance in the UI" idea this
+document originally deferred. Nothing currently gates on it.
+
+The other half of why the old rule bit harder than intended: the asset panel's single
+"Save changes" button sends `name`, `description` and `summary` together in one request
+whenever any one of them changed, so editing just the name also re-sent the other two
+fields' current values — which the API cannot distinguish from a deliberate edit, and so
+stamped `"human"` on fields the person had never touched, including ones still blank.
+That collateral marking is fixed alongside this change; provenance now reflects only the
+field actually edited.
+
+`name` remains the one field the suggestion rule (accept/reject, never a direct write)
+applies to regardless of provenance, because it is never blank rather than because of
+anything provenance records.
 
 ---
 
@@ -355,14 +374,13 @@ backend/app/
    Rejections are kept rather than deleted, so a re-run does not propose a tag the user
    already declined. Accepting a title goes through `apply_metadata` — the *human* path —
    because the user read it and chose it, which also stops a later run replacing it.
-7. `field_provenance` enforcement — **mostly done**, arriving with step 4 because
-   `summarize` was the first AI write path and shipping it without this would have meant
-   shipping the bug the rule exists to prevent. `services/assets.py::apply_ai_metadata`
-   reads the column, writes fields with no entry, skips ones marked `"human"`, and
-   returns what it actually wrote. No new provenance value was needed; an absent entry
-   already means "no person has touched this". What remains: applying it to `describe`'s
-   write path when that lands, and surfacing provenance in the UI so a user can see which
-   fields the AI wrote.
+7. `field_provenance` enforcement — **landed with step 4, then reverted post-M6.**
+   `apply_ai_metadata` originally read the column and skipped a field marked `"human"`.
+   In practice that made "Generate"/"Summarize"/"Describe" silently do nothing on a
+   field the person had often never actually typed into — see "What `field_provenance`
+   records" above for why, and for what it does now instead: every explicit press
+   overwrites unconditionally, and the column is kept only as a record of who wrote a
+   field, not as a gate.
 8. ~~Bulk enrichment over a selection — including the `SelectionBar` embed deferred from
    M5.~~ **Done.** One `bulk_enrich` job over the whole selection rather than one job per
    asset: cancellation is per row, so N jobs would mean N Cancel clicks and N
diff --git a/docs/plan-of-attack.md b/docs/plan-of-attack.md
index a1aed58..c9681a4 100644
--- a/docs/plan-of-attack.md
+++ b/docs/plan-of-attack.md
@@ -137,8 +137,11 @@ class Asset(SQLModel, table=True):
 
 Two deliberate deviations from the SRS schema:
 
-- **`field_provenance`** implements FR 8.1.3 ("manual edits are never overwritten by a
-  later AI run without confirmation"). Without it that requirement has no mechanism.
+- **`field_provenance`** was built for FR 8.1.3 ("manual edits are never overwritten by a
+  later AI run without confirmation"), but no longer enforces it — see `docs/m6-ai-
+  enrichment.md`'s "What `field_provenance` records" for why a silent skip turned out to
+  be the wrong default. The column still records `"human"` vs `"ai"` per field; nothing
+  currently gates on it.
 - **Transcripts are rows, not a nested JSON blob.** `TranscriptSegment(id, asset_id,
   user_id, idx, text, start_time, end_time, words JSON)`. FR 10.1.4 requires returning
   *a matching snippet with its timestamp*; that means each segment must be individually
diff --git a/frontend/src/components/AssetDetail.tsx b/frontend/src/components/AssetDetail.tsx
index ad15411..ffb5d82 100644
--- a/frontend/src/components/AssetDetail.tsx
+++ b/frontend/src/components/AssetDetail.tsx
@@ -12,7 +12,7 @@ import {
   Wand2,
   X,
 } from 'lucide-react'
-import type { Asset } from '@/api/assets'
+import type { Asset, AssetUpdate } from '@/api/assets'
 import { tagsApi } from '@/api/tags'
 import { enrichmentApi } from '@/api/enrichment'
 import { formatCost, usageApi, type UsageTotals } from '@/api/usage'
@@ -199,11 +199,16 @@ export default function AssetDetail({
     if (!dirty || !name.trim()) return
     setSaving(true)
     try {
-      await update(asset.id, {
-        name: name.trim(),
-        description: description || null,
-        summary: summary || null,
-      })
+      // Only the fields that actually changed: the API marks every key it receives as
+      // human-written provenance (FR 8.1.3 bookkeeping), so sending name/description/
+      // summary as a fixed trio would stamp the two you didn't touch right alongside
+      // the one you did — including while they're still empty.
+      const changes: AssetUpdate = {}
+      if (name.trim() !== asset.name) changes.name = name.trim()
+      if (description !== (asset.description ?? '')) changes.description = description || null
+      if (summary !== (asset.summary ?? '')) changes.summary = summary || null
+
+      await update(asset.id, changes)
     } catch {
       // The store restores the server's version and surfaces the message; the panel
       // stays open so the edit is not lost.
diff --git a/frontend/src/components/EnrichmentButton.tsx b/frontend/src/components/EnrichmentButton.tsx
index 0a617a4..6534022 100644
--- a/frontend/src/components/EnrichmentButton.tsx
+++ b/frontend/src/components/EnrichmentButton.tsx
@@ -94,8 +94,7 @@ export default function EnrichmentButton({
           .
         </p>
       )}
-      {/* A finished job's own message, which is where "you wrote this yourself" and any
-          provider failure arrive. */}
+      {/* A finished job's own message, which is where a provider failure arrives. */}
       {!running && job?.status === 'error' && job.error_message && (
         <p className={STATUS_ERROR_CLASS}>{job.error_message}</p>
       )}

From 97d8ddc85b12e43491519747afdf3b42195a192c Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Fri, 18 Sep 2026 00:56:43 +0000
Subject: [PATCH 2/2] Fix Prettier formatting in AssetDetail.tsx

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FdoHBNLsz2cXKBVzsqKPoV
---
 frontend/src/components/AssetDetail.tsx | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/frontend/src/components/AssetDetail.tsx b/frontend/src/components/AssetDetail.tsx
index ffb5d82..012d835 100644
--- a/frontend/src/components/AssetDetail.tsx
+++ b/frontend/src/components/AssetDetail.tsx
@@ -205,7 +205,8 @@ export default function AssetDetail({
       // the one you did — including while they're still empty.
       const changes: AssetUpdate = {}
       if (name.trim() !== asset.name) changes.name = name.trim()
-      if (description !== (asset.description ?? '')) changes.description = description || null
+      if (description !== (asset.description ?? ''))
+        changes.description = description || null
       if (summary !== (asset.summary ?? '')) changes.summary = summary || null
 
       await update(asset.id, changes)