Umbra open SAR archive

Composite 2-3 passes of one site into a change image. Pass the STAC URLs of two or three acquisitions of the *same* site in time order. Colors encode change: bright green = signal appeared after the first date, magenta = signal vanished, grey = unchanged. Refuses to mix polarizations (HH vs VV are not comparable). Returns the image block plus a caption naming the color semantics and attribution -- and, when a pass carries no polarization metadata so same-polarization could not be *verified*, a structured caution text block naming that gap.

Return a SAR-literate reading of the scene at ``url``. The picture and the metadata card are produced deterministically; a model only interprets them. How that model is reached depends on whether this process holds a vision key: - **No key** (the public community host): this tool does **not** call a vision model. It returns the PNG plus a reading kit (SAR rules, the metadata card, the JSON schema). Produce the JSON, then call :func:`stamp_description` — that is the validation boundary (provenance-stamped ``{summary, observed_features, confidence, caveats}``). Until you stamp it, your prose is not pipeline output. A baked preview is typically 512 px; the kit names the size and that a local server can stream a higher-resolution GeoTIFF. - **Key configured** (local stdio): sends the picture to that vision model, validates through ``parse_description``, and returns the stamped object directly. The packaged SAR rules are the value over just looking at ``quicklook``: brightness is backscatter not sunlight, speckle is not structure, a dark patch may be radar shadow not water, one frame is not change. ``preview`` defaults to ``"auto"``: a baked catalog preview if this server has one, else a fresh render. ``"render"`` is refused on the public host.

Download one asset of an item to local disk, gated by a size check. SAR products can be multiple GB, so this is a two-step handshake: call it first with ``confirm=False`` (the default) to get the byte size without downloading, then again with ``confirm=True`` to fetch it. On a local or stdio server, ``confirm=True`` returns ``{asset, path, bytes}``. ``confirm=False`` returns a size-and-hint object and does not write a file. On a public host (``umbra serve --public``, COG streaming off), ``confirm=True`` does not transfer the file. It returns a link plan: ``mode`` ``"link"``, ``downloaded`` false, ``asset``, ``filename`` (a basename), the direct public ``href``, ``bytes``, and ``how_to_fetch``. When the href is https on the public open-data bucket, ``how_to_fetch`` is ``curl -L -C - -o <filename> <href>`` and ``umbra download ITEM --asset ASSET --dest DIR``. ``bytes`` is the STAC ``file:size`` or ``size`` when that is a non-negative integer; otherwise the host HEADs that href for ``Content-Length``, with no redirects and no retries. A href that is not on that bucket is not requested: ``how_to_fetch.curl`` and ``how_to_fetch.umbra`` are null, ``bytes`` is null, and ``note`` says the commands were omitted. A 404 or 410 from an allowlisted HEAD raises. Any other HEAD failure, including a full quick-request slot, leaves ``bytes`` null and still returns the plan. The hosted server never transfers the file. A local ``uvx --from 'umbra-py[mcp]' umbra-mcp`` server downloads to your own disk. ``confirm=False`` on this host returns the same size-and-hint object and does not write a file. It HEADs for ``Content-Length`` only when the href is https on the public open-data bucket, with no retries, and it does not follow redirects. A 404 or 410 raises, and so does any other non-redirect HEAD error. A redirect leaves ``bytes`` null. A href outside the bucket is not requested; ``bytes`` is the STAC ``file:size`` or ``size`` when that is a non-negative integer, otherwise null. A local server still HEADs the href and follows redirects. On a local server, ``dest_dir`` is unchanged unless ``UMBRA_MCP_DOWNLOAD_ROOT`` is set. When that variable is set, ``dest_dir`` must resolve inside it. The argument itself is unchanged. ``url`` is a STAC JSON URL, or a product filename (``2025-05-17-17-43-46_UMBRA-08_GEC.tif``) or acquisition stem (``2025-05-17-17-43-46_UMBRA-08``). A filename or stem is resolved through the catalog index the same way as ``get_item``. The hosted server already has that index. A local server needs ``umbra index fetch``. ``asset`` selects the file. All of these work: - a product type, any case: ``GEC``, ``gec``, ``SICD`` - a schema v2 key, including a collision suffix: ``gec``, ``gec_2``, ``csi``, ``sidd``, ``csi_sidd``, ``sicd``, ``cphd``, ``metadata``, ``stac_metadata``, ``di_gif``, ``file``, ``file_tif`` - the original filename (the asset's ``umbra-py:filename``) ``GEC`` and ``gec`` are the primary file. ``gec_2`` is the second GEC, not the primary. A name that matches nothing raises with the item's v2 keys listed (``gec, sicd, cphd``, …), not the uppercase product types.

Rank sites imaged more than once, before a change tool is pointed at one. Use this when the site is not already known. Each site's ``hrefs`` are that site's STAC URLs, oldest first, ready for ``pick_change_interval``, ``change_composite``, or ``stack_stats``. Prefer this, or ``search_catalog(area=..., fuzzy=True)``, over ``semantic=True`` when the name shares words with a task. No model is called. Argument reference (ranking, recency, cadence, baseline, return fields): https://umbra-py.space/find-repeat-sites/ Filters bound which acquisitions are ranked. ``bbox`` is WGS84 ``[min_lon, min_lat, max_lon, max_lat]``. ``intersects`` is one GeoJSON Point, or a Polygon or MultiPolygon (or a Feature, FeatureCollection, or GeometryCollection wrapping those). A point matches footprints that contain it. It cannot be combined with ``bbox`` or ``place``. ``place`` is geocoded only when ``bbox`` is omitted. ``area`` is a task-name substring; ``fuzzy=True`` is a deterministic loose match. When ``area`` matches no task and ``bbox``, ``intersects``, and ``place`` are omitted, it is geocoded. The result then adds ``matched_by`` (``geocode``) plus the resolved place and bbox. A name Nominatim does not know, and a geocoder that is busy or unreachable, set ``hint`` and return no sites (``resolved_area`` is the ``area`` that was tried). A task-name hit adds nothing. ``start`` / ``end`` bound which passes enter the pool (ISO date, bare year or month, or a relative expression). ``products`` is any of GEC, CSI, SIDD, SICD, CPHD. ``polarizations``, ``min_incidence``, ``max_incidence``, and ``max_resolution`` (metres) drop passes missing that field. ``top`` (default 20) is how many sites to return. ``min_passes`` (default 2) is the minimum depth: comparable passes when ``rank_by`` is ``comparable``, otherwise raw dated passes. ``rank_by`` is ``passes`` (default), ``comparable`` (largest same-polarization subset), ``recency`` (newest pass), ``span`` (first to last), or ``cadence`` (tightest median gap). For the three temporal orders, ``min_passes`` still counts raw passes. ``active_since`` / ``active_before`` select sites by newest pass; ``first_since`` / ``first_before`` select by earliest pass. Those keep each survivor's full history, unlike ``start`` / ``end``, which truncate the pool. ``max_revisit_days`` caps the worst gap and ``median_revisit_days`` caps the median gap. ``min_span_days`` / ``max_span_days`` bound the first-to-last span. A non-positive cadence or span raises ``ValueError``. With ``rank_by="comparable"`` the cadence and span filters use that subset. A bare year or month on a ``*_before`` bound includes the whole period. ``limit`` (default 500) sizes only a live or Canopy pool. A local index ranks each task over the whole index and ignores ``limit``. ``local`` matches ``search_catalog``: unset uses the index when present. A Canopy token (``$UMBRA_CANOPY_TOKEN``) draws the pool from the commercial archive, and ``local=True`` cannot be combined with it.

Find archived scenes that **look like** the acquisition at ``url``. This is the one search that lives in the pixels, not the metadata: where ``search_catalog`` filters by date/bbox/task-name, this renders the query item's quicklook, embeds it, and ranks the pre-embedded scene index by visual (cosine) similarity — "find scenes that look like this flooded field / crowded berth / runway". The query item is excluded from its own results. Requires a scene-embedding index built ahead of time with ``umbra embed build`` (a sidecar ``catalog.embed.db`` beside the catalog index) and a user-supplied multimodal embedding key (``OPENAI_API_KEY``, optionally ``OPENAI_BASE_URL`` / ``UMBRA_SCENE_EMBED_MODEL``) — the only model call is turning the query image into a vector; ranking is deterministic. ``asset`` picks which product's quicklook to embed (default ``GEC``; match how the index was built); ``model`` must match the index's embedding model. Returns the ranked matches as compact records (each with the acquisition's ``href``, so hand it to ``get_item`` / ``quicklook`` / ``change_composite`` next) plus the attribution line.

Find archived scenes matching a plain-language ``query`` over their imagery. Text-to-scene search: embeds ``query`` ("ships at a berth", "a flooded field") and ranks the stored *image* vectors by cosine similarity. It only works when the index was built and this query is embedded with the **same** joint CLIP-family model whose text and image encoders share a vector space (a dimension mismatch is reported as an error; ``model`` selects it). Requires the same prebuilt ``catalog.embed.db`` (``umbra embed build``) and embedding key as ``find_similar``; the only model call is turning the text query into a vector. Returns the ranked matches as compact records (each with the acquisition's ``href`` for ``get_item`` / ``quicklook``) plus attribution.

Resolve a free-text place name to a bounding box via Nominatim. Returns ``{"bbox": [min_lon, min_lat, max_lon, max_lat], "display_name": ...}`` so the agent can inspect the box before searching, or pass it straight to ``search_catalog``'s ``bbox``.

Return the full context card for one item. ``url`` is the JSON URL of an open-data STAC sidecar, a GEC/SICD (or other product) filename such as ``2025-05-17-17-43-46_UMBRA-08_GEC.tif``, or the acquisition stem ``2025-05-17-17-43-46_UMBRA-08``. A filename or stem is resolved through the local catalog index: the stem is matched against asset hrefs, which is what distinguishes ``sar-data/tasks/``, ``sar-data/task-data/``, and ``open-data/`` — the name itself does not. Zero matches is an error. More than one match lists the candidate STAC URLs; pass one of those. The hosted server resolves a filename or stem with the catalog index it already loaded. A local server does the same after ``umbra index fetch``. When the index contains a STAC URL, the indexed snapshot is returned — the same document ``search_catalog`` already answered from — rather than a live re-fetch that can disagree on asset filenames or product types. When the server is configured with a token (``$UMBRA_CANOPY_TOKEN``) and ``url`` is a bare acquisition id (no ``://``, and not a filename or stem), it is instead looked up in Umbra's authenticated commercial archive by a keyed STAC search — the retrieval complement to ``search_catalog`` over the paid catalog, since archive cards carry an ``id`` (not a fetchable sidecar URL) as their hand-off pointer. A filename or stem still resolves from the local open-data index when a token is set. On the public host a URL is fetched only when it is https on Umbra's public open-data bucket. Any other URL is refused before a request. A bare id that is not in the local index is refused the same way. The card (:meth:`UmbraItem.to_llm_context`) carries the id, ISO datetime, place, bbox, resolution, polarization *with the change-detection caveat*, the per-product-type explanations and asset URLs, and the mandatory attribution line — everything a model needs to reason about the scene.

Report on the local catalog index, if one has been built or fetched. Lets a long-running agent decide whether searches will be instant (index present) or pay a live S3 walk, and how stale the snapshot is. ``end`` is the newest acquisition that is not ``datetime_suspect``. ``datetime_suspect_count`` is how many kept rows carry that flag. ``quarantined_future_count`` is how many source rows the last build dropped for a datetime more than a day ahead (they are not in the index). ``unique_captures`` is how many rows are the canonical copy of their collect; it equals ``items`` until republished copies have been linked. ``path`` is the index file locally and ``null`` on the public host, which does not reveal the server filesystem.

Return a grounded reading of *what changed* between passes. Sibling of :func:`describe_scene`. The composite and the per-block dB grid are produced deterministically; a model only narrates change those numbers support. Pass the STAC URLs in time order; mixed polarizations are refused. This tool **streams Umbra COGs** to build the composite, so a public community host refuses it (use a local ``uvx --from 'umbra-py[mcp]' umbra-mcp`` server, or describe each pass with :func:`describe_scene` / :func:`quicklook` on baked previews). On a local server: - **No vision key:** returns the composite as an MCP image block plus a reading kit (SAR change primer, the dB grid, the JSON schema). **You** (the client's model) narrate, grounded in the numbers. - **Key configured:** sends picture + grid to that vision model, validates through ``parse_narration``, and returns a provenance-stamped ``{summary, changes[], confidence, caveats[], change_stats}`` object.

Scan a site's whole series and name the two passes worth narrating. The deterministic first half of the "watch a site over time" story, and the step that turns a wall of passes into an answer. Given the STAC URLs of a site's passes (2 or more, any order), it co-registers them into a ``to_stack`` cube, reduces it with ``stack_stats``, and returns the *one* consecutive pass-pair whose measured change stands **furthest clear of the speckle detection floor** — the pair a reader should look at first. No model is called: a number picks the frames (the determinism boundary applied to frame selection), so the choice is reproducible and safe to quote. Why this exists as its own tool: ``narrate_change`` reads a *pair* of passes, but a site has many, and "which two?" is a measurement question, not a model one — a picture past three dates encodes nothing a model can separate, so the pair has to be chosen before the model is called. This answers it. The returned ``urls`` are the two chosen passes in time order, ready to hand straight to ``narrate_change`` (or ``change_composite`` to see them): the two tools chain into scan → narrate with no server and no model in the selection. Returns ``selected_interval`` (the interval's ``earlier_id`` / ``later_id``, their datetimes, the ``changed_fraction`` and signed ``mean_delta_db``, the cube's ``false_alarm_fraction`` speckle floor, the ``excess`` of the change over that floor, and ``stands_clear`` — whether the change cleared the floor by a usable margin), ``urls`` (the earlier then the later pass's STAC URL), and the license/attribution. When ``stands_clear`` is false the series' own largest change is still inside the speckle, so the pair is offered but the reader is warned the difference may be interference rather than change. ``selected_interval`` is ``None`` (and ``urls`` empty) when the series has fewer than two comparable passes or no interval reported any change — there is no pair to narrate. ``crs="utm"`` (the default) grids the cube in the site's UTM zone so the change fractions weigh equal ground; ``max_size`` and ``change_threshold_db`` match ``stack_stats``. Refuses to mix polarizations (HH vs VV are not comparable). Requires the ``load`` extra.

Return a scene's quicklook PNG as an image block. Prefers a baked catalog preview (``preview="auto"``, the default) so a public host can show the radar scene without proxying Umbra COGs. **A baked preview is typically 512 px** (the published weekly sidecar). ``max_size`` applies only to a fresh ``preview="render"`` stream; when a cached thumbnail is smaller than requested, the caption says ``max_size`` was ignored *and* includes a JSON ``{"image": {source, width, height, requested_max_size, substituted}}`` sidecar. For a higher-resolution picture (typically 1024 px), run a local ``uvx --from 'umbra-py[mcp]' umbra-mcp`` server, or download the GEC href from ``get_item``. ``preview="baked"`` refuses rather than stream; ``preview="render"`` always streams (needs ``[viz]``, refused on ``umbra serve --public``).

Search Umbra's catalog and return compact context cards. Filters (all optional, combine freely): ``bbox`` as ``[min_lon, min_lat, max_lon, max_lat]`` in WGS84 degrees; ``intersects`` as a GeoJSON object (a ``Point``, ``Polygon``, or ``MultiPolygon``, or a ``Feature`` / ``FeatureCollection`` wrapping one; a point matches footprints that contain it) to keep only footprints intersecting it -- a tighter spatial filter than the rectangular ``bbox``, and mutually exclusive with ``bbox`` / ``place``; ``place`` as a free-text name geocoded to a bbox; ``area`` as a substring of the Umbra task (site) name (set ``fuzzy`` to match it loosely -- word-order- and punctuation-independent and typo-tolerant, resolved deterministically with no model call, so ``"beet piler"`` still reaches ``"Beet Piler - ND"`` and ``"utah centerfield"`` still reaches ``"Centerfield, Utah"``). When ``area`` matches no task and ``bbox``, ``intersects``, and ``place`` are omitted, it is geocoded and that footprint is searched; the result then adds ``matched_by`` (``geocode``), ``resolved_place``, and ``resolved_bbox``. A name Nominatim does not know sets ``hint`` to ``no task named X; try place= or bbox=``. When geocoding is busy (the Nominatim gate, or the public host's quick outbound pool) or unreachable, ``hint`` is ``no task named X; geocoding was unavailable`` and ``resolved_area`` is the ``area`` that was tried. A task-name hit adds nothing. Set ``semantic`` instead to treat ``area`` as a plain-language *description* of a site whose name you don't know and resolve it to the closest task names by *meaning* via the prebuilt embedding index, so ``"grain storage north dakota"`` reaches ``"Beet Piler - ND"`` -- a query that shares no word with the label, which ``fuzzy`` cannot and should not fake; ``semantic`` and ``fuzzy`` are mutually exclusive, and the resolved task name is returned as ``resolved_area`` with the ranked candidates in ``semantic_matches`` so you can audit and retry with a different one, and ``min_score`` -- a cosine threshold in ``[0, 1]`` -- drops weak aliases, so a description that matches nothing confidently returns an empty result rather than an arbitrary top pick); ``start``/``end`` as acquisition-date bounds -- an ISO ``YYYY-MM-DD`` date, a bare year/month (``2024``, ``2024-03``), or a relative expression (``today``, ``yesterday``, ``3 months ago``, ``last month``), resolved deterministically with no model call; ``products`` to restrict to product types (any of GEC, SICD, SIDD, CPHD); the SAR acquisition-property filters ``polarizations`` (keep items exposing any of these, e.g. ``["VV"]`` -- the filter that keeps a change comparison like-with-like), ``min_incidence`` / ``max_incidence`` (view incidence-angle bounds in degrees) and ``max_resolution`` (keep items at least this fine, in metres) -- each a hard predicate that excludes items missing that metadata; ``canonical`` to keep the preferred copy of a collect Umbra republished under another prefix (``True``), only the other copies (``False``), or every item when omitted; ``limit`` to cap results (on the public host a value above 100 is clamped to 100 and the result includes ``note``; the argument and its default are unchanged); ``max_per_task`` to cap items per site (use 1 for a one-pin-per-site overview). ``local`` selects the backend: leave it unset to use the on-disk index when present (instant) and fall back to a live S3 walk otherwise. When the server is configured with a Canopy token (``$UMBRA_CANOPY_TOKEN``), the search runs against Umbra's authenticated commercial archive instead and ``source`` is reported as ``"canopy-archive"``. Returns the per-item context cards, not full STAC JSON, to protect the context window; call ``get_item`` for one item's full metadata. Prefer ``fuzzy=True`` (or :func:`find_repeat_sites`) when ``area`` shares words with a task name; that path is deterministic and needs no key. ``semantic=True`` is only for a description that shares **no** tokens with the label. It requires a task-embedding index built ahead of time with ``umbra semantic build`` and a user-supplied embedding key (``OPENAI_API_KEY``, optionally ``OPENAI_BASE_URL`` / ``UMBRA_EMBED_MODEL``); it raises a helpful setup error otherwise, so it never runs implicitly. The only model call is turning ``area`` into a vector -- the resolved search over the chosen backend stays fully deterministic.

Check whether a site's passes are one measurement, before measuring them. ``stack_stats`` refuses a selection whose rasters were made by *different* ``umbra convert`` settings — a calibrated pass differenced against an uncalibrated one puts the difference between two conversions on the time axis and reports it as change on the ground. This is the question that refusal answers, asked ahead of it: it reads each acquisition's ``UMBRA_*`` conversion record straight from its raster header (kilobytes per pass, no pixels, no co-registration) and groups the selection by what its pixel values *are*. Returns ``agrees`` (whether ``stack_stats`` would accept the selection), ``groups`` largest-first — each with the conversion ``record`` its members share, their ``item_ids`` and the ``hrefs`` to re-run on — plus ``shared`` (the record they all agree on) when they agree, ``refusal`` (verbatim the error ``stack_stats`` would raise) when they do not, and ``unreadable`` for any source that could not be opened at all, which is an incomplete answer rather than a mixed one. Reach for it in two situations. **After** a ``stack_stats`` call fails with a provenance refusal: ``groups[0]["hrefs"]`` is the largest set of passes that *do* agree, so re-running ``stack_stats`` on those is a measurement rather than a guess — the refusal itself cannot name that subset. And **before** measuring a long series you did not assemble yourself, since the check costs a header read per pass and saves the whole cube when the answer is no. A series of Umbra's published products as they ship (the common case) agrees trivially: every pass reads as ``"(unrecorded)"``, which is one value, so ``agrees`` is true and ``shared`` is empty. **No model is called.**

Measure how much a site changed across its passes, as numbers. The quantitative counterpart to ``change_composite`` / ``timescan``: those return a picture of *where* change is, this returns the *how much*. Pass the STAC URLs of a site's passes (2 or more, any order); they are co-registered onto one shared grid and reduced to a per-pass statistics series plus a net baseline-to-latest record — mean/median/spread of each pass, the signed decibel change against the pass before it, the fraction of ground that moved past ``change_threshold_db``, and that fraction as an **area in km²**. Change is always reported in decibels (a ratio of backscatter is a difference on the log scale). The default ``crs="utm"`` puts the grid in the site's UTM zone so every cell covers the same ground and the area figures mean something; pass ``crs="EPSG:4326"`` for a lon/lat grid, and the areas come back ``None`` rather than wrong. ``extent="intersection"`` (the default) compares only ground every pass covers; ``"union"`` keeps all of it and compares per cell where two passes overlap. ``bbox`` (lon/lat ``[min_lon, min_lat, max_lon, max_lat]``) narrows the measurement to a sub-area, e.g. one facility inside the scene. ``blocks=N`` adds the *where* to the *how much*: the scene is cut into an N x N grid (6 is a good default; 0, the default, skips it) and the extra ``spatial`` key reports each block's own net change, a compass label and lon/lat centre to locate it by, and the consecutive pair of passes that block moved most between — plus ``peak_block`` and an ASCII ``grid_text`` heat-map of the net change. Ask for it whenever the question is *where on this site* something happened, or when a scene-wide mean looks small but the picture from ``change_composite`` shows an obvious hot spot: a change confined to one corner is averaged away scene-wide and stands out per block. ``block_series=True`` (which needs ``blocks``) keeps each block's *whole* pass-to-pass sequence rather than only its peak interval, so a block that drifted a little every pass reads differently from one that jumped once and held. It is the largest payload this tool emits — ``blocks`` × ``blocks`` × (passes − 1) records — so ask for it only when the *shape* of the history is the question, not the size of the change. ``speckle_filter="boxcar"`` or ``"lee"`` averages **speckle** down before anything is measured. Speckle is not sensor noise: it is the interference pattern coherent illumination makes on rough ground, and on single-look imagery it scatters as widely as its own mean — so on a quiet site an unfiltered per-cell decibel delta is mostly interference rather than change. ``"boxcar"`` averages every window (most variance removed, blind to the edges it averages across); ``"lee"`` averages only where a window is no more variable than speckle alone explains, so edges and bright points survive. Reach for it when the change you are looking for is small, or when the numbers look noisy against a picture that looks quiet; leave it off when the question is about *fine* structure, because what it spends is resolution — a ``speckle_window``-cell window (odd, 5 by default) makes each cell report ground that many cells across. The returned ``caveats`` say so, so a quoted figure carries the trade with it. Refuses to mix polarizations (HH and VV are not comparable). **No model is called** — this is arithmetic over streamed COG overviews, so the numbers are reproducible and safe to quote. Read the ``caveats`` before interpreting them: Umbra's open products are not radiometrically calibrated, so a decibel value is relative to the same ground on another date, and look geometry moves backscatter too.

Validate a client-produced scene reading and stamp provenance. This is the interpretation boundary for the no-key kit path: you read the picture ``describe_scene`` returned, produce ``{summary, observed_features, confidence, caveats}``, and pass it here. The library checks the shape, stamps CC-BY and the AI-interpretation note (the model never sets those), and — when ``image`` is the record from the kit — appends a deterministic caveat if the picture was a smaller baked preview. Until this returns, the reading is unvalidated prose, not pipeline output.

Validate a client-produced change narration and stamp provenance. Sibling of :func:`stamp_description` for the no-key ``narrate_change`` kit. Pass the JSON you produced plus the ``change_stats`` grid from the kit so the stamped object carries the numbers the prose was grounded in.

Summarize a whole time series into one activity image. Pass the STAC URLs of a site's passes; the timescan encodes per-pixel temporal statistics so bright/colored areas are *where activity happened* over the series. Returns the image block plus attribution.

Report only the acquisitions **new** since the last check of this site. SAR's value for monitoring is its cadence, so the natural workflow is *standing*: run the same search each time and act only on what changed. This tool packages that idempotent delta — call it in one turn, come back a day (or a session) later and call it again, and it returns just the passes published in between, not the whole list again. The scheduler is you: ask the agent to check the site, and it reports the delta. Define the site with the same filters as ``search_catalog`` — ``place`` (geocoded to a bbox), ``area`` (task/site-name substring, loosened by ``fuzzy``), ``bbox``, ``products``, and ``start``/``end`` date bounds. The watch is identified by a stable ``name`` derived from those filters (pass an explicit ``name`` to run several independent watches over the same site); ``reset=True`` re-establishes the baseline, reporting everything as new. State persists in the local catalog index's ``meta`` table (created on first use), so a watch survives across sessions with no extra setup. ``local`` selects the search backend exactly as in ``search_catalog`` — leave it unset to use the on-disk index when present and a live S3 walk otherwise; for true monitoring a live walk (``local=false``) catches freshly published passes. **No model is called** — this is pure set arithmetic over the deterministic search — and the returned ``new_items`` are context cards ready to hand straight to ``change_composite`` / ``timescan``, closing the standing-analyst loop (new pass → composite → describe) without leaving the conversation. On the public host (``umbra serve --public``, COG streaming off) nothing is stored. Every call reports the current matches as new, says so in ``note``, and caps ``new_items`` at 50 (``new_total`` is the full count). The watch name is capped at 128 characters. A local server still persists the baseline in the index.