Colormeris

Colormeris is a set of static web pages that turn colors in scientific figures back into numbers. You load a PDF, PNG or JPG and click to mark the colorbar. The page reads values out of the colors and exports them as CSV. It can also save a project zip, which you can load again later to review or re-run the extraction.

There are three pages:

• index.html is the landing page.
• extract.html turns figure colors into numbers. It has three tools, which share the loaded file, viewer, colorbar calibration and project:
  • Heatmap (extract.html#heatmap) gives one value per cell of a gridded heatmap.
  • ROI (extract.html#roi) measures signal inside regions of interest drawn on an image, such as an IVIS luminescence image.
  • Map (extract.html#map) reads near-continuous images, such as spectroscopy maps with many bins or fluorescence images, as a dense value matrix with axis coordinates, plus line profiles.
• colormaps.html browses, compares and identifies colormaps (see Colormaps).

In extract.html, switch tools at any time with the Heatmap | ROI | Map control in the top bar. The file, the PDF page and each tool's panels stay. One project zip holds the work of every tool.

Everything runs in the browser. Files are never uploaded anywhere, except when you run the heatmap agent: it sends page images and extracted values to OpenRouter and the models you pick.

Heatmap tool

1. Open a PDF, PNG or JPG. You can use Open file…, drag and drop, or paste an image from the clipboard. For a PDF, choose the page and the render resolution. The default is 216 dpi.
2. Panels: add one panel for each heatmap in the figure. Each panel has its own grid, labels and colorbar. In a PDF, panels belong to the page they were made on. Switching pages keeps them and shows the new page's own panels. The Panels card links to other pages that already have panels.
3. Grid: press Place grid corners (G). Click the outer top-left corner of the heatmap, then the outer bottom-right corner. Colormeris guesses the number of rows and columns from where the colors change and fills them in. Correct them if needed, or press Detect to guess again. Then paste the row and column labels, or open Row and column labels and press Read labels with AI: a vision LLM (OpenAI GPT-6.1 Sol by default; change it in the Model field below the button) reads them from the current view for this panel only. Zoom so the grid and its labels are visible. It uses the Agent's OpenRouter key or proxy, costs about a cent, and Undo reverts it. You can drag the corner handles to adjust. Uncheck Keep rectangular for skewed or rotated scans. Dot plots (e.g. scRNA-seq), which have no cell borders and dots of different sizes: set Click on to Dot centers before placing, then click the centers of the top-left and bottom-right dots and type the rows and columns (Detect needs cell borders). The grid is drawn through the dot centers. Each dot is sampled inside a circle of 70 % of its own radius, found from where its color meets the background; a cell without a dot is left blank. Shape (Square / Circle) and Sampled area are under Options in both modes.
4. Colorbar: press Place colorbar ends (B) and click both ends of the bar along its middle. Lines within 3° of vertical or horizontal snap straight; hold Alt to turn snapping off. Next, click at least two labelled ticks on the bar and type their values (T). Values between ticks are interpolated linearly, and values beyond the outer ticks are extrapolated. If the tick labels are raw numbers on a logarithmic bar, choose Log₁₀. If the tick values are uneven along the bar (for example a binned bar labelled 1e4, 2e4, 2e5, 5e5 at its steps), choose Piecewise, or Piecewise log10 for positive values over decades: values are then interpolated between neighbouring ticks only. A binned bar (flat colors with steps between them) is found by itself: each cell gets the middle value of its bin, and a bin past the outermost tick gets that tick's value as a bound (at most or at least). Cells whose color is not on the colormap, such as gray "ND" cells or a black diagonal, get no value. Cells past an end of the bar (paler or darker in the bar's own colors) read as that end. Drag an end handle to adjust the bar: under When dragging an end, ticks, choose whether the ticks move with the bar (keeping their proportions) or stay in place on their printed marks. Drag the line itself to move the whole calibration, ticks included. To place a tick exactly, type its position in % along the bar (0 at the start, 100 at the end) in the tick table. Under the sampled strip, the card suggests the Matplotlib, CMasher, Crameri, cmocean, colorcet, seaborn, CarbonPlan, NCL, SciVisColor, CARTOColors, MATLAB or R packages colormap the bar matches (for example jet or viridis_r, with its ΔE2000 score and a link to the colormap viewer). With two ticks the direction follows the values; this works in the ROI tool too. Known colormap. When the figure's colorbar is missing, tiny or badly compressed, set Colors to Known colormap and pick the map with the same search field as the colormap viewer (any of the 1063 maps; type name_r or tick Reversed for the reversed map). Two ticks appear at 0% and 100% of the colormap, valued 0 and 1 (1 and 10 on a log scale): type the figure's values at its ends, or keep them for values relative to the colormap. Type a tick adds more by value and position in % along the colormap, and every tick's position can be edited in the table. Placing the colorbar ends is then optional; if you do, ticks clicked on the bar count too. When a figure's bar matches a known map, the suggestion under the strip has Use its colors to switch to that map's exact colors. This works the same in the ROI and Map tools.
5. Results: hover a table cell to find it on the image. Turn on Reconstruct to repaint each sampled area with the color the matched value predicts, so you can check the match by eye. Cells whose color is far from every colorbar color (ΔE above the threshold) are outlined in red. A cell whose color is not on the colormap (a gray "ND" cell, a black diagonal) shows "–" and its reason in the tooltip. Values that are only bounds (past an end of the bar, or in an open end bin) start with ≤ or ≥.
6. Export:
   • Download CSV saves the matrix (rows × columns).
   • Long CSV saves one row per cell, with the sampled RGB, ΔE, a status (ok, flagged, clamped, off-map, empty), a bound (<= or >= for values that are only bounds) and a reason for off-map cells (gray, hue, far).
   • Export project (.zip) saves everything needed to reload the project.

Navigation: scroll to zoom and drag to pan. Space-drag or middle-drag always pans. F fits the image to the view, Esc cancels the current tool, and Ctrl/⌘+Z undoes.

ROI tool

1. Open the figure as for the heatmap tool. Panels work the same way, including across PDF pages.
2. Grid (optional): if the animals are shown in boxes, place the grid. Include the name labels above each row so every animal sits at the same place in its box. Check rows and columns (Detect guesses them) and type the box names in reading order, e.g. B-a11, B-a16, …. Remove drops the grid.
3. Colorbar: place the ends and ticks as for heatmaps. For bars with a ×10ⁿ multiplier, type values like 1.4e9.
4. Regions:
   • Pick Ellipse (E), Rectangle (R) or Polygon (P).
   • Drag to draw an ellipse or rectangle; hold Shift for a circle or square.
   • For a polygon, click its points, then click the first point, double-click or press Enter. Backspace removes the last point.
   • With Copy new regions into every grid box on, a region drawn inside a box is copied into every box. Copies share their shape: dragging a handle on any copy resizes all of them.
   • Dragging the inside of a copy nudges only that box's copy. Alt-drag moves all copies, and Reset nudges puts them back.
   • Regions drawn with the option off, or without a grid, are single regions. They report under the box they sit in.
   • Different shapes can be mixed freely. Select a region to rename or delete it (Del).
5. Scale (optional): Measure scale bar, click both ends of a known distance, then enter its length in cm, mm, µm or nm. Areas are then also reported in that unit squared.
6. Results: pick a measurement for the table (boxes × regions). Download CSV saves every measurement for every region copy. Turn on Show signal to tint the pixels counted as signal.

Measurements for each region copy:

Column	Meaning
area_px	pixels inside the region
signal_px	pixels with colormap color (not gray)
sum	total of the colorbar values over the region (gray = 0), like total flux in pixel units
mean	sum / area_px, like average radiance
mean_signal	sum / signal_px
max	highest value
flagged_px	colored pixels that match no colorbar color (ΔE above the limit), e.g. colored text. They are still counted as signal with their nearest colorbar value.
area_<unit>2, signal_area_<unit>2, sum_x_area	with a scale bar: areas in real units, and sum × pixel area

Background: a pixel whose CIELAB chroma is at or below the gray threshold is the photograph, i.e. no signal. The default is 20, which removes the JPEG color noise seen in published IVIS figures while keeping the dimmest overlay colors. Adjust it under Settings → Matching and check with Show signal.

Map tool

For figures whose pixels are the data: spectroscopy maps with so many bins that they look continuous, or false-color fluorescence images with a calibration bar.

1. Open the figure as for the heatmap tool. Panels work the same way.
2. Plot area: press Place plot area (G) and click the top-left and bottom-right corners of the image area, inside the axes.
3. Colorbar: place the ends and ticks as for heatmaps.
4. Scale (optional), for images with a scale bar but no axes: Measure scale bar, click both ends of the scale bar, then enter its length and unit (cm, mm, µm or nm). Profiles, the matrix and the status bar can then report positions in that unit.
5. Axes (optional): press Add x ticks (X), click two or more labelled ticks on the x axis and type their values. Do the same for y (Y). Ticks stay on the plot area's edges (x on the bottom, y on the left), so only their position along the axis counts; drag them along the edge to adjust, or type a position in % across the plot area in the tick table (0 at the left or top, 100 at the right or bottom), for example for a tick exactly at the plot's corner. Each axis can be linear or Log₁₀. Without ticks, coordinates are pixels from the plot's top-left corner.
6. Results: Bin size 1 gives one value per pixel. Larger bins give one value per N × N pixels, the median of each channel, which smooths out JPEG noise. Hover the image for the value and axis coordinates under the pointer. Reconstruct repaints the plot with the matched values. Show flags marks colors far from the colorbar in red (ΔE above the threshold) and colors at the top or bottom end of the colorbar in magenta or cyan.
7. Profiles: press Draw profile (L) and click both ends of a line, for example a spectrum at one delay or a line scan across a cell. X axis picks what the plot's horizontal axis shows: Axis (the calibrated axis the line mostly runs along), px (distance from the line's start in pixels) or the scale-bar unit (distance from the line's start in real length, for example µm). It is the same setting as Coordinates in the Results card. When a choice is not available it falls back to the axis, then the scale bar, then pixels. Length sets how long the line is, in the same units; the start and direction stay. Width averages ± pixels across the line. Shift- or Cmd-click profiles (chips or lines) to overlay several in the plot, each in its own color; Select all overlays every profile. Scale plots the values as they are, so signal levels compare directly, or divides each profile by its own maximum or mean, so their shapes compare. Hover the plot or a line on the image to trace it: both show the same sample, with its position and value. Drag the ends or the line to move it; hold Shift to move it only horizontally or vertically. Hold Ctrl (⌘ on a Mac) when you start dragging a line to drag a copy and leave the original in place. Ctrl+C / Ctrl+V (⌘ on a Mac) copy and paste the selected profiles, also into another panel or page: there a copy keeps its place within the plot area, so the same line scan lands in the same spot of a panel of another size. () moves the edited profile across the plot area, perpendicular to itself, and back at each edge, so the plot plays like an animation. sets how fast, in page pixels per second. With the line keeps going until no part of it is left in the plot area, which covers more of a slanted line's corners. Parts of any profile outside the plot area are left out of the plot and the CSV. The value axis stays fixed during the sweep, so heights compare between frames. leaves the line where it is; one undo puts it back where it started.

What the numbers can and cannot tell you:

• Compression. JPEG stores color at half resolution and in 8 × 8 blocks. Single pixels can be off by several percent of the range; the summary and Show flags show where. Bins of 4–8 px bring the worst case down a lot. On a synthetic viridis map saved at JPEG quality 0.6, the worst pixel was off by 10% of the range at bin 1 and by 2.4% at bin 8; the mean error was 0.6%.
• Levels. A colorbar can only tell so many values apart. The summary reports how many distinct colors the sampled bar has, at most 256 for an 8-bit colormap and fewer for a short bar.
• Display units. Published fluorescence images are almost always contrast-adjusted. Values are in the units of the colorbar as displayed, which supports comparisons within the image, not absolute intensities.
• Clipping. Pixels at the top or bottom color of the bar may be saturated: their true value can lie beyond the bar. The summary counts them.

Large maps take a moment: a 600 × 400 px map at bin 1 takes about 1–2 s with CIEDE2000, and CIE76 (in Settings) is about 10× faster.

Heatmap agent

The Agent card in the heatmap tool calibrates every heatmap on the chosen pages by itself:

1. Connect. Open Settings → Agent and enter your OpenRouter key. Or keep it out of the browser: put OPENROUTER_API_KEY=… in .env (ignored by git), run npm run proxy, and set API base URL to http://localhost:8787/api/v1 with the key field empty.
2. Pick a mode with the toggle at the top right of the card:
   • Agent only: the LLM does everything. It looks at page images with pixel rulers, zooms in, places the grid, labels, colorbar and ticks, and checks them with overlays.
   • Agent decision: the LLM reads the page and describes each heatmap. A decision model (Decision model, default typesafe/jev-1.14) locates its cells and colorbar, one heatmap per request. Colormeris snaps those boxes to the pixels and counts the rows and columns itself. The LLM then reads the numbered tick labels and the row and column labels from one image per heatmap. When it is unclear which colorbar belongs to a heatmap (several colorbars, or one far away), the LLM picks it from lettered candidates, or copies another panel's colorbar when several blocks share one. Strips of cells with their own colorbar (such as a -log10(p) row) are extracted as panels of their own. This takes fewer LLM steps and images: on the example PDF's page 5 (10 heatmaps), 3 steps and about $0.19, of which the decision model cost half a cent.
3. Pick models. The LLM needs image input and tool calling; the default is anthropic/claude-sonnet-5.5. Requests use prompt caching, and the card shows the share of prompt tokens read from the cache. Then choose the pages and press Run agent.
4. Review. Checks that an external agent answered below Min. confidence (through the API) appear under Needs review. A rejected panel gets a red dot and stays there until you accept it or it changes. Add a note on what is wrong and press Redo with agent to have the agent fix it. Delete panel removes a wrong extraction outright (undo brings it back).

The agent retries a failing tool at most 3 times per panel, then skips that step and reports it. It stops after 8 failed calls in a row or after Max. steps. The log shows every step and the cost so far.

Page images and extracted values are sent to OpenRouter and the model providers you pick.

Settings

Settings (top right) holds what belongs to you rather than to a figure. It is saved in the browser as you change it.

• Agent: OpenRouter key (stored only if Remember is ticked), Min. confidence, Max. steps, API base URL, and the prompts of both agent modes and the label reader.
• Matching: color difference, ΔE flag limit and, for ROI, the gray threshold. New panels start with these values; changing one applies it to every panel of that tool in the open project (Undo reverts it). Each panel keeps its own copy in the project zip, so a project reopens with the values it was made with.
• Hotkeys, in sections (General, View and pages, Calibration, ROI, Map): press Change and then the new key. A key taken from another action of the same tool moves to the new one; ROI and Map actions can share keys, since only one tool is active. Esc, Space, Enter, Backspace and Delete are fixed. Keys are shown for your computer: Ctrl on Windows and Linux is ⌘ on a Mac, so the same settings work on both. Control (⌃) on a Mac and the Windows key are left to the system.
• Colors: the colors of marks drawn over figures (grid and plot area, colorbar, flags, highlight, scale bar, ROI regions and masks, map axis ticks, profiles, trace and flags). Change them when they blend into a figure's colormap; Reset restores a default.
• Import and export: Export settings writes colormeris-settings.zip (one settings.json); Import settings reads it back, or a bare settings.json. The API key is left out unless you tick Include the API key; importing a file without a key keeps the current one.

Agent API

extract.html exposes typed actions as window.colormeris, so any program can drive the tools without clicking pixels:

await colormeris.run('set_grid', { topLeft: { x: 40, y: 60 }, bottomRight: { x: 520, y: 400 } });
await colormeris.run('add_tick', { at: { x: 560, y: 400 }, value: 0 });
await colormeris.run('get_results');   // → { ok, result } or { ok: false, error }

docs/agent.md is the full reference: every action and argument, typed questions and reviews, the heatmap agent's tools, limits and system prompt. It is generated from the code with node scripts/agent-docs.mjs, and a test fails when it is out of date.

Colormaps

colormaps.html is a viewer for 1063 colormaps: the 87 of Matplotlib, 53 from CMasher (cmr.amber), 36 from Crameri's Scientific colour maps (cmc.batlow) 21 from cmocean (cmo.thermal) 80 from colorcet (cet_fire) and 12 from seaborn (rocket, vlag, and the qualitative deep to colorblind), named as matplotlib registers them, 58 from CarbonPlan (carbonplan.fire_light, carbonplan.fire_dark: one version for light and one for dark backgrounds; rainbow and sinebow are the same in both, so only _light is listed), 162 of NCAR's NCL color tables (ncl.BlueRed; stepped tables such as ncl.precip_11lev are shown as steps, as matplotlib draws them, and NCL has no types, so their groups are ours), 107 from SciVisColor (sciviz.yg1; non-sequential ones sorted by eye into diverging, multi-sequential, outlier-range and discrete), and 34 from CARTOColors (carto.Burg) and 2 from MATLAB (matlab.parula, matlab.lines; MATLAB's other maps, such as jet and hot, are Matplotlib's) and 411 from R packages, as listed by R Charts from the collection (, , , , and others from ggpomological, ggthemr, tidyquant, tvthemes, vapoRwave, colorBlindness, dichromat and cartography; R Charts shows continuous palettes as 30 samples, which are interpolated linearly, and discrete ones as steps; the groups are ours, sorted by lightness; RColorBrewer and viridis are already Matplotlib's, copies of maps listed earlier are left out, and so is , which has 51 colors). CARTOColors are palettes of up to 7 steps; their continuous versions interpolate the 7 colors in sRGB. Maps with the same colors as one already listed are left out: the shifted cyclic CMasher maps (), Crameri's categorical maps (), berlin, managua and vanimo, which ship with Matplotlib, and , which is . colorcet registers most maps three times (, , ); only the short name is kept, and its rotated cyclic maps (, ) and 256-color Glasbey palettes are left out, as are seaborn's 6-color palettes (), which are subsets of the 10-color ones, and NCL's copies of other maps (, , an older ), test tables and 170-class tables (). in can show the maps of one source; the Python packages and the R packages are each a group of sources, and sub-headings read or . Every sub-heading shows how many maps it holds. Within each sub-heading, maps are ordered by similarity, so families sit together. Each row has a reverse button (⇄). In sequential sections, maps that run the other way than most of their neighbors start reversed (the button is then on), so a section reads one way; under ⋯ still flips everything. It has five tabs: , , , and . The tab, the open map and the comparison are kept in the page link, so any view can be shared.

• Browse shows each map as vector strips: as seen, as simulated for protanopia, deuteranopia and tritanopia, and in grayscale. Under Views, pick one view to show only that strip, wider; the page remembers your choice.
• Hover a strip to see the hex, RGB, position and lightness (L*) of that color. Click to copy the hex. With the keyboard, focus a strip and use the arrow keys (Shift moves by 10).
• Four columns rate each map as uniform (U), CVD-safe, grayscale-safe and readable: ✓ yes, ~ partly, ✕ no. Hover a rating for the numbers, and press ? in the header for how they are made. Click a column name to sort by it (best first, then worst first). On phones the ratings are pills under the name.
• Readable is about reading values back from colors, as Colormeris does. A color can be off by a few ΔE in a figure, so a difference under 3 ΔE2000 is not trusted. A map is less readable where values 5% apart look the same (flat zones), or where a color has a look-alike at least 10% away (ambiguous). jet is readable but not uniform.
• The search box filters by name. Filters keeps maps rated yes (All four ✓ for maps that pass everything). Under ⋯ are Reversed, which flips every map, and the full sort menu (for example Most linear L*). Each group folds; Rainbow and Others start folded and open when a search matches them.
• Click a name, or the arrow at the end of its row, to see one map in detail. On wide screens it opens in a panel at the right and the list stays in place; on narrower ones it opens under the row. Esc closes it. It starts with one sentence per rating and four small square plots (in one row when there is room, else 2 × 2): L*, ΔE2000 steps, chroma and hue (hover a plot title's ⓘ for what to look for). Folded below are the L* stats with a table per view (smallest difference between values at least 10% apart, distinguishable levels, flat and ambiguous share; hover a row to mark the two closest colors on its strip), where values get confused on the strip, and the map's references. Expand all opens every fold at once.
• Press + on a row to add it to the comparison (up to 10). A tray at the bottom lists the chosen maps. Compare shows them side by side: strips, four plots in a row (L*, ΔE2000 steps, chroma and hue) with all maps on the same axes and one legend (hover a name to highlight that map in every plot), and a sortable table of the key numbers (All numbers shows the rest; the best value in each column is bold). You can reorder or remove maps there. Use Copy link to share it (colormaps.html?compare=viridis,cividis#compare).
• Identify: load an image (choose, drop or paste), then drag along its colorbar. The page snaps the line to the bar, samples it and lists the closest maps and their direction. matches the colors of a heatmap without a colorbar, but cannot tell the direction. opens a match in , and takes the image and the line to . The image never leaves the browser.

How heatmap values are computed

• Cell color: each channel's median over the central part of the cell. The Sampled area setting controls how much, 50% by default. Using the center avoids grid lines, anti-aliased edges and JPEG noise.
• Colorbar: sampled at 256 evenly spaced points from one end to the other. Each point averages ±half-width pixels across the bar.
• Matching: the cell color is converted to CIELAB and matched to the nearest colorbar sample by CIEDE2000 (CIE76 is optional). The match is refined between neighbouring samples. That position along the bar is then converted to a value using the ticks.

Project zip layout

project.json            per panel: tool (heatmap, roi or map), page, grid, colorbar,
                        labels, settings, review (accepted/rejected), for ROI
                        the regions and scale bar, and for maps the bin size,
                        axis ticks and profiles
README.txt
source/<original file>  the uploaded PDF/image
source/page-<n>.png     the rendered image of each page that has panels
data/<panel>.csv        heatmap: matrix of values
data/<panel>_long.csv   heatmap: per-cell values with page, RGB and ΔE
data/<panel>_rois.csv   ROI: measurements per region copy
data/<panel>_map.csv    map: matrix of values with x and y at bin centres
data/<panel>_map_long.csv  map: one row per value with ΔE and flags
data/<panel>_profile_<name>.csv  map: values along each profile
agent/actions.json      changes made through the agent API, in order
agent/decisions.json    typed decisions (answer, confidence, source, applied)

The agent logs are written but not read back when a zip is reopened; panel reviews are (they live in project.json). Each panel records its page. Its coordinates are pixels in that page's source/page-<n>.png. Regions copied into every box are stored relative to a box, where a box spans 0–1 in each direction, with per-box nudges. A zip opens in either tool with everything in it. Version 1 zips load too; one without a tool is a heatmap project. Zips saved before the ROI tool was renamed (tool ivis) are refused with an error rather than opened as heatmaps.

Development

More documentation: docs/agent.md (agent reference, generated), docs/research.md (prior work, novelty, experiment plan), and CLAUDE.md (architecture, conventions, status and next steps for Claude sessions).

The app is written in TypeScript (strict) and built with Vite. extract.html and colormaps.html each load one module entry from src/pages/; index.html is static. pdf.js, JSZip and the OpenRouter TypeScript SDK (loaded only when the agent runs) come from npm. Node 22.18 or newer:

npm install
npm run dev       # dev server on http://localhost:8000 with hot reload
npm run check     # lint (Biome), typecheck (tsc) and unit tests (Vitest); run before every commit
npm run build     # production build in dist/ (npm run preview serves it)
npm run format    # format with Biome
npm run proxy     # optional: OpenRouter proxy on :8787 using the key in .env
npm run docs      # regenerate docs/agent.md after changing the agent
npm run fixtures  # regenerate the synthetic calibration heatmaps in src/testing/fixtures/

CI (.github/workflows/ci.yml) runs lint, typecheck, tests and the build on every push and pull request. Pushes to main that pass CI are deployed to GitHub Pages by .github/workflows/pages.yml (repository settings → Pages → Source: GitHub Actions). docs/typescript.md has the coding conventions.

The code is grouped by feature. Files without DOM code have unit tests next to them (grid.ts and grid.test.ts).

Folder	Purpose
src/pages/	one entry per page: extract.ts starts the extraction app, colormaps.ts the colormap viewer
src/types/	shared domain types: project, panels, grids, colorbars, results, settings, workspace, agent, colormaps
src/core/	shared by both apps: color spaces and ΔE (color.ts), CVD simulation (cvd.ts), grid geometry and cell sampling (grid.ts), colorbar calibration (colorbar.ts), snapping rough boxes to cells, colorbars and tick labels (locate.ts), the colormap library (colormap-library.ts, data in src/generated/) and matching a bar to a known colormap (colormap-match.ts)
src/extract/	extract.html: project model and files (project/), settings (settings/), PDF loading, the canvas viewer and the scale bar
src/extract/workspace/	the shell the tools share: workspace.ts (state, modes, tool switching) and one file each for undo, overlay drawing, pointer input, the Grid, Colorbar and Panels cards, opening files and pages, export and hotkeys
src/extract/heatmap/, roi/, map/	the three tools: pure extraction (sampling.ts, geometry.ts and quantify.ts, field.ts) and the tool UI
src/agent/	agent API and heatmap agent: action schema (schema.ts), window.colormeris (api.ts), prompt, tools and retry guard (llm.ts), decision mode (decide.ts, decision-tools.ts), the loop (runner.ts); the Agent card and label reader in ui/
src/colormaps/	colormaps.html: viewer.ts (state, tabs, page) with one file per tab in tabs/, plus strips, plots, footer, URL routing (route.ts), ratings (metrics.ts), references and recoloring
src/styles/	base.css for every page, plus , and