TODO / Roadmap

Pending work for aggregate_api, newest thinking at the bottom of each list. What has landed is in CHANGELOG.md and the git log; completed plans move to dev/done/.

Items oddly from the aggregate list

[Session-Build-Clobbers-The-Trailer]: a session build(...) of a program that shadows a library name replaces the library entry's note{} and tags{}.

Landed

[Short-Links] DONE (1.43.0), plan dev/done/plan-short-links-and-cards.md. Copy link copies agg.mynl.com/x/<code>, 5 characters for the entry and 3 for the tab, hashed and never stored; long links still work. /x/<code> returns a link preview card (name, note, social-card.png) and the site has a generic one. Still owed, the author's: deploy, then opengraph.xyz, LinkedIn Post Inspector and Facebook Sharing Debugger on a short link and on the bare site.
[API-Editable-Aggregate-Source] DONE (1.42.0), plan dev/done/plan-api-editable-aggregate-source.md. aggregate installs editable from ../../worktrees/aggregate_REFACTOR through [tool.uv.sources], so uv sync keeps the worktree library; the dependency is aggregate[elt]>=1.61.0 with no ceiling, and tests/test_private_names.py guards the two private names. Still owed, the author's: the VPS symlink ~/worktrees/aggregate_REFACTOR before the next refresh.sh, and the two skill edits drafted in the plan's log.
[Follow-Aggregate-1.61] DONE (1.41.0), no plan: the two follow-ups aggregate 1.56.0 and 1.59.0 name for the API. The strip grades each clause of a reinsured verdict, so a clean net view is green, and an unmatched bracket is the library's parse error with its position.
[Multi-Statement-Programs] DONE (1.40.0), plan dev/done/plan-multi-statement-programs.md. The box takes several statements, the statements before the last are filed and only the last is built, and the API refuses what build() refuses, for every client. Controls that transform the box refuse several statements in the status strip. Left open: Sharpen, Hints and Cede act on the last built object when one statement was edited and not built; POST /v1/decl/parse still parses each statement against the default base; Reformat keeping references needs a library plan (option [c] in the plan).
[Chfagg-Aggregate-Parity] DONE (1.38.0), no plan: the consumer half of aggregate 1.35.0 to 1.41.0. A chfagg takes the arithmetic form, Evaluate and Sharpen run on it, with the More Sharpen leaf and the bs_window and tail_behavior exhibits, and the word list follows 1.41.0. The floor landed at 1.42.0, [API-Editable-Aggregate-Source].
[Chfagg-Kind] DONE (1.37.0), no plan: the library's chfagg law (aggregate 1.30.0 to 1.34.0) builds, prices and exports, and the word list follows 1.34.0. Owed upstream: _KIND_WRITE_ORDER writes agg before the chfagg it references, so the library's own .agg export does not reload.
[Status-Bar-Text-Answers] DONE (1.36.0), plan dev/done/plan-status-bar-text-answers.md. A conversion statement (tweedie, stable) answers as text in the aLL strip and as the cell value in Excel and Sheets; the word list follows aggregate 1.23.0 and ylt became yelt. Still owed, the author's: npm run push for the Sheets template.
[Eltq-Version] DONE (1.35.1), no plan: About, /v1/meta and the status page list the eltq version.
[Loss-Table-File-Gate] DONE (1.35.0), plan dev/done/plan-loss-table-file-gate.md. A loss table file path is read for private requests only: the request text, references through the preview, and the object cache. Still owed, the author's: deploy to the VPS, and 192.168.4.0/24 in AGGAPI_PRIVATE_CIDRS on DOOB's launcher. Optional later: a library switch that refuses file sources at the read, as a second layer.
[ELT-Keywords-And-Eltq] DONE (1.34.0), no plan: the consumer half of the library's loss tables, aggregate 1.12.0 to 1.18.0. eltq>=2.13 as a direct dependency and the word list regenerated against 1.18.0. aggregate[elt] and the floor landed at 1.42.0, [API-Editable-Aggregate-Source].
[Menu-Preferences] DONE (1.33.0), no plan: tire kicking on a phone. Tables, Perspective and Editor collapse under a Preferences row after Apps, and the menu scrolls within the screen. Still owed, the author's: the check on a real phone.
[Greater-Tables-8] DONE (1.32.0), plan dev/done/plan-greater-tables-8.md, the consumer half of the library's [Format-Sheet-Block-Filter] (aggregate 1.10.0, on PyPI as 1.11.0). Floors greater-tables>=8.1.0 and aggregate>=1.11.0; no greater_tables warning in any response. Still owed, the author's: the plugin's re-lock to greater-tables 8.1.0, uncommitted in aggregate-relativity, and the by-hand acceptance against the restarted server.
[Excel-Library-And-Frames] DONE (1.22.0 to 1.26.0), plan dev/done/plan-excel-library-and-frames.md, five steps, one bump each: [lab-all-rows-switch] 1.22.0 (All rows as a menu item and Ctrl+Shift+A, the density grids window, nine seconds to one), [readable-arguments] 1.23.0 (#AGG! for every refusal, p on [0, 1]), [frames-route] 1.24.0 (GET /frames, AGG.FRAMES, AGG.FRAMEURL), [library-functions] 1.25.0 (AGG.LIBRARY, AGG.PROGRAM), [pane-frames] 1.26.0 (the pane's Whole frames section). Still owed, the author's: the Excel acceptance list in desktop Excel (npm stop, npm start, then the plan's four items), the Ctrl+Shift+A check in Edge and Firefox, and the deploy. Upstream ask for csv-grid: refresh the status line after _applyHeight.
[Excel-Program-Ranges] DONE (1.21.0), no plan: tire kicking after 1.20.0. The three handle-taking functions accept a program as a range, a refused program returns #DECL! and the reason as cell text, and the development build calls the local 19456 server.
[Fast-Frames] DONE (1.17.1 to 1.20.0), plan dev/done/plan-fast-frames.md: [orjson-frames] 1.17.1, [gzip-level] 1.17.2, [parquet-route] 1.18.0, [lab-all-rows] 1.19.0, [excel-decl-arguments] 1.20.0. Full density frames for the Lab (All rows, 3.5M cell budget), Excel (Power Query over Parquet) and R. Still owed, the author's: the Excel acceptance list in desktop Excel (Power Query and byte stream split among them), the host sync with --extra parquet, and the Caddy check that a .parquet body is not re-encoded. Findings: the 8.5 s All rows took on 65,536 by 12 was the grid rendering every row unwindowed, not the parse, fixed in 1.22.0 ([Excel-Library-And-Frames]); the exeqa_ display binning (right edge, not mass-weighted) is still open; the R client is its own plan, next.
[Excel-Hosting] DONE (1.17.0), plan dev/done/plan-excel-hosting.md. /excel serves the add-in built by scripts/build-excel.ps1 (or .sh), no-cache, with a production manifest under https://agg.mynl.com/excel. On the host, done: the Caddy limits (verified 2026-10-18), the deploy, and AGGAPI_CACHE_MAX=100 (2026-10-20, in the systemd unit). Still owed, the author's: the plan's acceptance list against agg.mynl.com (web in Edge and Chrome, Windows desktop, a version bump reaching a reopened workbook).
[Excel-Core] DONE (1.16.0), plan dev/done/plan-excel-core.md. The excel/ add-in: AGG.BUILD, AGG.EXHIBITS, AGG.EXHIBIT, AGG.Q and AGG.VERSION over a local server, with the recipe table, session id, memo and request limit. Still owed: the plan's acceptance list in desktop Excel (npm start), which nothing here could run. Later: [excel-pane], [excel-types], AGG.FRAME.
[Session-Parse] DONE (1.15.0), plan dev/done/plan-reformat-explodes-session-references.md. Reformat previews the program in the caller's fork and renders each statement's spec, so own declarations inline like library entries and a second statement may name the first. No web/ change.
[Server-Parsed-Key] DONE (1.14.0), plan dev/done/plan-cache-key-from-parsed-program.md. The object id hashes the parsed program, its references' stored specs, the grid, versions and the library digest, so a redefined name builds a fresh object and an id means the same thing in any process. Adds library_references_only for Excel. Still owed: an upstream ask for a lossless str on Distortion and Copula (see the plan's execution log, divergence 1).
[Warm-Library] DONE (1.13.4), plan dev/done/plan-1.13.4-warm-library-at-startup.md. create_app() reads the library before serving, so a fresh server's first page load can no longer fork a half-read base (Capstone.ExposureRating failing on Capstone.Gross). Still owed: the manual restart check in the plan. The library half landed in aggregate 1.5.2, dev/done/plan-1.5.2-thread-safe-load.md.
[Exact-Tail-Leaf] DONE (1.13.0), plan dev/done/plan-1.13.0-exact-tail-leaf.md, the api half of the library's [NTT-Exact-Tail-Chart] (aggregate 1.5.0). More ▸ Exact tail draws the nttagg's exact mass and survival as base-10 exponents beside the float FFT; the adapter learns log10_density and log10_probability. Still owed: the manual checks in the plan (browser draw time on Wilson2017.DiceOfDice among them), and a ruling on the two nttagg_gaps smoke failures its execution log describes.
[Deep-Link-View] DONE (1.10.0 to 1.11.0), plan dev/done/plan-1.10.0-deep-link-view.md, two steps. [view] (1.10.0): view=<group>:<leaf> lands a deep link's build on a named tab, keys or labels in any case, with a note in the strip when the view cannot be shown. [copy-link] (1.11.0): Menu ▸ Copy link writes a link to the entry and tab on screen, against AGGAPI_PUBLIC_URL (default https://agg.mynl.com/). Still owed, by eye in a visible window (the plan's execution log says why): the scroll anchor on a view link, tall and short, and the link copied clipboard path. Deploying to the public site is the author's.
[Version-Source] DONE (1.9.1), plan dev/done/plan-1.9.1-version-source.md, shared with aggregate 1.2.1, aggregate-relativity 0.10.0 and aggregate-ntt. Every version the server reports is read live from __version__, so a bump runs no uv sync.
[NTT-Version] DONE (1.9.0), plan dev/done/plan-1.9.0-ntt-version.md. About lists aggregate-ntt when the compiled engine is installed, from /v1/meta.ntt_version.
[Pentagon-Readout] DONE (1.8.0, tweak 1.8.1), plan dev/done/plan-1.6.0-pentagon-readout.md (numbered 1.6.0 when drafted, landed after the DecL help took 1.6.0 and 1.7.0). The PnL group's Pentagon notes moved from Bootstrap tooltips to one fixed readout in the figure's lower-left corner, driven by hover, focus and tap. Still owed: the iPad check on a real device.
[DecL-Help] DONE (1.6.0 to 1.7.0), plan dev/done/plan-1.7.0-decl-help.md, two steps, the api half of the library's [DecL-Help-Pages] (aggregate 1.2.0a15). [anatomy] (1.6.0): a ? above the ✕ in the program box opens the library's anatomy page in a popover over the dimmed output, with Load buttons that fill the editor without building; GET /v1/decl/help and /v1/meta.decl_help. [card] (1.7.0): Menu ▸ DecL on one page, the card on a sheet that prints to one Letter landscape page. Both one below the plan's numbers, since the pentagon readout plan had not landed. Still owed, outside the plan: the help reaches PyPI users only once aggregate 1.2.0 ships, and the pin floor moves only if the author wants it guaranteed.
[Example-Databases] DONE (1.4.0 to 1.5.0), plan dev/done/plan-1.4.0-example-databases.md, two steps. [databases] (1.4.0): Menu ▸ Examples from picks the shipped .agg file behind the Examples menu, per viewer, over GET /v1/examples?database=. [deep-link] (1.5.0): /?database=<stem>&example=<name> selects the file and builds the entry, for the paper's links, desktop and phone. Still owed, outside the plan: raise the pin floor to the release carrying aas-paper.agg and deploy. (The two nttagg tests library 1.2.0a13's [NTT-Grid-Cap] broke were fixed after 1.5.0, test only: request log2=6 for the dice, and a wider program for the binning case.)
[nttagg-Exact-Log] DONE (1.3.0), no plan doc: a tweak off the author's 1.2.0 tire-kicking, landing with a library reshape. p_exact_log10 joins the Density grid, and serializers.mass_columns stops the display binning summing it as though it were a mass. p_exact left density_df at library 1.2.0a7, so the exact rationals are no longer reachable through the api at all; see the note at the end of CHANGELOG.md 1.3.0 and the open item below.
[nttagg-Density-Column] DONE (1.2.0), no plan doc: a tweak off the author's 1.1.0 tire-kicking. More > Density asks for p_exact_float on an nttagg, through a per-kind addition to the curated column subset (DENSITY_COLS_BY_KIND in web/src/main.js) rather than by sending the kind down the whole-frame path. The reading is in the tails, where the float pipeline runs out of precision and the exact answer still has digits. No api change; the route already served the column by name. Detail in CHANGELOG.md 1.2.0.
[nttagg] DONE (1.1.0), plan dev/done/plan-1.1.0-nttagg.md. The library's exact bounded compound (aggregate 1.1.0's nttagg) is admitted, served and labeled: four kind touchpoints, the NTT Aggregate label in both copies of KIND_LABEL, the keyword in the editor's pool, and the exact rational column kept out of both display payloads while the CSV keeps it. capability.py deliberately untouched; the composability fix was the library's and landed in aggregate 1.2.0a3. Detail in CHANGELOG.md 1.1.0.
- _KIND_OF_CLASS, SUPPORTED_KINDS, _KIND_ORDER, BuildResponse.kind. The last is not cosmetic: a kind the route classifies and the response model does not list 500s after the build has run.
- p_exact out of the JSON density_df and the format=ir document, through one shared helper so the two display paths cannot drift. The headline case puts a 410-character rational in a cell, and the column also matches the p-prefix test that decides what a display grid sums.
- Tests gated on the parser rather than on a version string (tests.conftest.requires_nttagg), so the suite stays green on the released 1.0.1 the pin still admits.
- Two hard-coded kind lists in tests/test_examples.py widened, ungated. The second one was already red against the library's new ExactDiceOfDice entry; the plan had identified only the first.
- [NTT-Accuracy], carried forward to the backlog: the accuracy exhibit that would show what makes the kind interesting.
[Release-1.0.0] DONE (1.0.0), plan dev/done/plan-1.0.0-release.md. The first public release, timed with the library's 1.0.1 on 2026-10-07. Three packaging defects meant no installation route delivered a working app, and all three were invisible to both suites because both read the working tree in place. Detail in CHANGELOG.md 1.0.0.
- [tool.uv.sources] removed, aggregate>=1.0.1,<2 pinned. The table was fatal to uv pip install git+..., not merely unhelpful: uv honors it for a git-installed project and looked for the sibling worktree inside the repo.
- package-data carries static/assets/*. A package_data glob does not recurse, so the wheel shipped index.html and none of the thirteen files it loads: 200 for /, 404 for every asset.
- scripts/release-check.ps1, the rehearsal that found all of it. Builds, inspects the wheel, installs into a clean environment, runs the suite against the installed package, serves the app and fetches every asset.
- test_health_needs_no_thread_token stopped walking app.routes, which FastAPI 0.142 no longer flattens, so it passed on the locked pins and failed on the ones a fresh install resolves. tests/test_packaging.py added.
- PyPI metadata, Development Status :: 4 - Beta, README install section rewritten around pip install aggregate_api.
- Author's, not yet done: uv publish, git tag -a v1.0.0 (this repo has no tags at all), the GitHub release with the distributions attached. See V:\dev\RELEASE-PROCESS.md.
[Punch-List-Oct-07] DONE (1.0.0a198 to 1.0.0a200), plan dev/done/plan-a198-punch-list.md with its execution notes beside it. The author's 2026-10-07 punch list after kicking the a197 tires: six items landing as five version bumps across three repositories, all executed 2026-10-07.
- 1.0.0a198 [quick-re], the five faults in the Quick Re quote sheet: the table cycles with the page's table view off a table document, the minimum rate on line becomes the fourth candidate so binds names it, a row carries its layer name beside its span, a rate on line is written as a bare fraction and rounded to four figures so the clause and the sheet agree to the digit, and an occurrence tower stops at the policy limit. policy_limit joins the build response. Detail in CHANGELOG.md a198.
- 1.0.0a199 [narrative-bounds], white-space: pre-wrap on the narrative panes, and the Bounds form on a P&L seeding off the closing block rather than the gross one. net_premium joins the preview response and all three premium figures are named wherever the pair was. Detail in CHANGELOG.md a199.
- library 1.0.0a397 and plugin 0.9.0 [pnl-emphasis], PnL Summary and Waterfall bolding gross, each net of tier and the closing net, and waterfall_df's CR becoming Margin ratio. The plan named plugin 0.8.0, which [Waterfall-Capital-Catchup] had already taken, so it landed as 0.9.0.
- library 1.0.0a398 [reins-labels], the Reinsurance plot's annual triple named Gross, Ceded to occurrence, Net of occurrence. The document-wide legend strip goes from three chips to six, which the plan anticipated; the revert is one line if that reads worse in use.
- library 1.0.0a399 and api 1.0.0a200 [capstone], Capstone.PC replaced in place by Capstone.Capped, a capped quota share that peels top-down. Detail in CHANGELOG.md a200.
[PnL-Pentagon] DONE (1.0.0a195, 1.0.0a196), plan dev/done/plan-a195-pnl-pentagon.md: the PnL group gains a Pentagon leaf drawing gross, ceded and net as one figure, at a solvency level picked from a strip of six mini pentagons. a195 is the api half (the POST /v1/objects/{id}/pnl/pentagon route serving one frame per level, the can_pnl_pentagon flag, the runner in its own aggregate_api/pnl.py); a196 is the figure, ported from aggregate-asp, with its level strip and its note. Library a392 ([Delta-Row-Marked], the stitched impact row stops printing a negative standard deviation) and a393 ([PnL-Pentagon], PnL.pentagon_df). Not in this plan, after the author's ruling at review: the Waterfall leaf gaining the same strip, which needs the library's conditional capital bases parametrized off PERCENTILE_LADDER first and is its own upstream plan. See the plan's execution log for the five other divergences.
[PnL-Exhibits] DONE (1.0.0a194), plan dev/done/plan-a194-pnl-exhibits.md: the Overview group becomes uniformly the wrapped book and the PnL group uniformly the ledger, which gains a Tail leaf for the closing margin's ladder. PnL Summary draws one table read across a row, PnL Waterfall's second table is three capital bases with the cost of each (every capital figure the negative of a193's, reading as capital rather than as the margin in the bad state), and PnL Ledger inherits Skew. The P&L summary card loses its leaf; nothing it carried is unreachable. Library a388-a391 ([Overview-Engine], [PnL-Summary-One-Table], [Waterfall-Capital], [Ledger-Skew]). check-nav.mjs's EXPECTED table brought current, having fallen eleven findings behind the library. Detail in CHANGELOG.md a194 and the four library entries. Next: a195, the PnL pentagon, which depends on this bump's column rename.
[Tower-Generator] DONE (1.0.0a193), plan dev/done/plan-a193-quick-re-tower.md: the Quick Re row takes a second button, Tower, that lays one span out as a stack of quote-sheet-rounded layers with a climbing placement ladder, names them Occ 1 up, and cedes the stack as one clause. Reverses the row's own "ONE layer, deliberately" ruling, on the author's reversal of 2026-10-06. The arithmetic is new web/src/tower.js, ported from aggregate-asp except for the ladder, which aSP walks downward. The attachment box now reads a bare 0, which Add re gains too. SPA only, no library change. Detail in CHANGELOG.md a193.
[Decl-Parse] DONE (1.0.0a192), plan dev/done/plan-a192-decl-parse.md: POST /v1/decl/parse returns the (kind, name, spec) triple the parser understood, one entry per statement, building nothing and caching nothing. Requested by aggregate-asp for its [Program-Import] step, which reads the structure of a pasted program rather than regexing it out of pprogram. A non-finite float in a spec serializes as "inf" / "-inf" / "nan" through the new serializers.spec_to_payload, and a parse failure is the existing ErrorReport 422 plus a statement_index. Detail in CHANGELOG.md a192.
[Punch-List] DONE (1.0.0a191), the author's 2026-10-01 punch list: the rate menu items light (upstream pnl_program(premium_style=) shipped as library a386), the net-of-occurrence subtotal reaches PnL Summary, Waterfall and Lab Relativity (library a387, plugin 0.7.0), Insurer Summary drops the by-leg table, the Waterfall gains Margin ratio, Bounds auto-computes on a P&L at the ledger premium, and the Pricing Bounds heading names the base constraint. Detail in CHANGELOG.md a191 and the two library entries.
[Evaluate-Targets] DONE (1.0.0a190), author's tweak after kicking the a186 tires: Evaluate offers the full CoC | LR | Premium switch (it passed targets: ['premium'] and drew a bare label), a CoC or LR resolving its premium through the pentagon at the anchor before evaluating; a typed Premium stays gross of expenses, and the help is the author's sentence verbatim.
[Expenses] DONE (1.0.0a186), plan dev/plan-a182-pricing-pnl-forms.md part [expenses]: expense_ratio on the four pricing request bodies (a premium target reads as gross, the engine runs on its technical part, the arithmetic in the runners once), the preview's gross_premium / net_of_expense_premium pair, the shared expense box on Calibrate, Allocate and Evaluate, and the P&L ledger pair on the Bounds preview and the dormant Evaluate band. This closes the plan's four parts (a182, a183, a184, a186; a185 was the unrelated matrix fix); the plan stays in dev/ until the author calls it done. The rate menu items lit at a191, when the upstream premium_style ask shipped as library a386.
[Sentence-Forms] DONE (1.0.0a184), plan dev/plan-a182-pricing-pnl-forms.md part [sentence-forms]: the five pricing form mounts read as one sentence each in the Quick Re idiom, the band label and ◦ spacer retired, and a ? per form carries the explanation. Visual acceptance (all six states, phone wrap) awaits the author's browser pass.
[PnL-Dropdown] DONE (1.0.0a183), plan dev/plan-a182-pricing-pnl-forms.md part [pnl-dropdown]: the PnL split button with fixed labels (PnL, and a caret with PnL (rate), xPnL, xPnL (rate)), the stale-object reset bug fixed by deriving from the box, form and premium_style on the pnl route, and the can_pnl_rate / can_xpnl flags. The two rate items stayed dark until the upstream ask shipped (pnl_program(premium_style=), library a386); the api feature-detects the signature, so the a191 sync lit them with no change here.
[Evaluate-Auto] DONE (1.0.0a182), plan dev/plan-a182-pricing-pnl-forms.md part [evaluate-auto]: Pricing / Evaluate on a P&L computes on arrival and the form band goes dormant (drawn, dead, explained) rather than hiding its field pairs. The plan's remaining parts ([pnl-dropdown], [sentence-forms], [expenses]) land as a183 onward; the plan stays in dev/ until the author calls it done.
[Matrix-Legibility] DONE (1.0.0a181), plan dev/plan-a179-lab-tab.md section 9: the matrix is a custom series with pixel band gaps and its own height, row names wrap, an empty axis label draws no name, and an ordinal xy axis reads ChartAxis.categories and draws as a trend. From the author's review of the first Lab render.
[Matrix-Renderer] DONE (1.0.0a180), plan dev/plan-a179-lab-tab.md section 8: the ECharts adapter reads chart IR version 4 and draws the matrix panel kind. Required for any build running library a379 or later, since the adapter's version check is one-sided and an unraised pin blanks every chart.
[Lab-Tab] DONE (1.0.0a179), plan dev/plan-a179-lab-tab.md: third-party charts and exhibits appear under a new Lab tab, discovered through aggregate.plugins entry points. AGGAPI_PLUGINS_ENABLED / AGGAPI_PLUGINS_ALLOW, the manifest on GET /v1/meta, a dynamic seventh nav group, one generic lab:* loader. capability.py deliberately untouched; see the plan's section 5. Verified against aggregate-relativity 0.1.0, which is not a dependency of this repo and must not become one.

Next up: three things, and then we are in good shape

The author's plan, set out 2026-07-31 (Friday), to finish over the weekend. Each becomes its own dev/plan-*.md and its own version bump.

a) One payload per table, and move the switch — done (a34, a35)

The connector arrived while this was being planned. GT 1.8.0 shipped irToGridInput(doc), 1.9.0 renamed the package back to greater_tables, and both landed: a34 the rename, a35 the single fetch. Every table under 500 rows is now switchable off one document, the densities stay on the bulk path, and the stats_df divergence that a33 introduced is fixed and guarded.

The switch did not move. It is still in the header menu, deliberately: where it belongs depends on (c), so settle that first. The argument recorded last time still stands, and is now cheaper to act on: the flip costs no round trip, which is a point in favor of putting it back next to the table.

Left over from this work:

Look at the tables. Still nobody has seen one in the page, and there is now more to look at: 1.7.x spanner rules, and an interactive view whose numbers arrive as raw values with IR-declared formats rather than pre-rounded off the wire. Build a portfolio, then walk Overview, More, Reins and Price flipping the switch on each.
Upstream, reported: include_raw='data' and irToGridInput do not compose. 'data' covers numeric, date and bool columns only, so a string data column gets no raw value and the adapter then throws on the whole document. Three Price frames carry one. We pass the explicit column list instead. Written up in c:/s/ai/greatest-tables/dev/note-from-aggregate-api-include-raw.md.
Two harnesses in dev/scripts/: check-frames.py (every frame, every kind, both routes agree) and check-adapter.py (a document is interchangeable with its FrameResponse). Run both when greater-tables moves. A third joined at a39: check-exhibits.py (every exhibit, every kind: capability and envelope agree, byte determinism, insurer blocks match the frame routes row for row). A fourth at a44, and it is the app side rather than the server: capture-capability.py writes one build response per kind to dev/fixtures/capability.json, then node dev/scripts/check-nav.mjs runs the real rules from web/src/nav.js over them and prints a leaf-by-kind grid. Run it whenever a leaf, a gate or a library registration moves.

Exhibits endpoint, remaining slices (a39 landed the routes)

The [Exhibits-App-Endpoint] phase of the library's dev/plan-exhibits.md landed its additive slice at a39: the capability and envelope routes plus check-exhibits.py. Still open, gated on the library's PnL insurer framing review: the menu-from-capability rewrite (applyKindGating and the hardcoded NA tables read the capability response; has_reins gating folds in; chips gray out, never hide; unknown-to-the-page exhibits list under More), deleting the migrated ROW_FLAGS / FORMATS entries and the main.js title and caption literals (pricing FORMATS were to stay; as of 2026-08-12 they delete too, per dev/plan-pricing-exhibits.md), wiring check-exhibits.py into CI, and extending capture_fixtures.py to the envelopes. (The name collision that blocked the last of those is settled: a62 renamed the chart fixtures to dev/fixtures/charts.json, so exhibits.json is free for the library's exhibit envelopes.) The [Exhibits-App-Cleanup] decision (retire _drop_raw_moments by re-pointing the frame routes at exhibit raw and insurer, or keep the frame routes as the raw CSV path forever) is taken during the endpoint phase proper.

Settled at a71: every table the library publishes is drawn as published. Twelve of the nineteen table leaves take the exhibit route, which is all of them that can. ROW_FLAGS is empty and FORMATS holds only frames no exhibit serves. The [Exhibits-App-Cleanup] decision above is therefore taken: the migrated entries are deleted, and the frame routes stay as the raw / CSV api path, rendered by inference.

What is left, and none of it is a published table: Bounds (stays with the library's dev/plan-exhibit-official-channels.md phase 10, the one open item of that plan), and the two densities, bulk and the one exception the author scoped out. More Sharpen closed at a94: loadSharpenAudit became loadExhibitLeaf on the a255 sharpen exhibit, the two block ledes went with it, and tables.FORMATS is empty. Pricing closed at a83 to a85 via dev/plan-pricing-exhibits.md (three subtabs served by three exhibits on result objects, [Pricing-Keyed-On-Result]). The asks were written up as aggregate_REFACTOR/dev/note-from-aggregate-api-round-6.md; items 1, 2, 3, 5 and 6 closed upstream at a251 to a256, and item 4 closed through the pricing plan.

b) Consolidate the graphs on ECharts, and fix the options

Most of the way (a36, a37). a36 took the Plotly path out: engine.js, plotly-panels.js, the draw control, the dependency and the 535 kB chunk. twoPanelData stayed.

a37 did item 5 of the a30 graph notes, the horizontal y-to-x readout, plus three things the first browser look turned up: full x widening only one panel, the return period capped at 1e5, and densities drawing as pyramids rather than steps.

Superseded, 2026-08-09, by dev/plan-plot-ir-api.md. Configuring ECharts turned out to be the smaller half. The author's ruling is that the app assumes IR-based charts always, and purist about it, so the open items from the a30 graph notes moved into that plan (item 7) alongside the chart-rendering items from dev/api-punchlist.md, and they are done as part of the migration rather than against the app's own builders. dev/graphs.md was deleted once its content landed there. The plan also carries the deletion of the server-rendered matplotlib route and the adoption of the four converted emitters.

Most of that plan landed at a62: the 2-D adapter, all seven documents adopted, the purist availability rule, and the control strip off the document. exhibits.js is gone. What is left of it is item 7's remaining punch items (7.1 zoom disorientation, 7.2 double-click reset, 7.4 the readout in the legend) and the heatmap declaration upstream.

Every chart on the page is a different picture now and nobody has seen one. Ten documents replay clean through the real adapter in node (dev/scripts/smoke-charts.mjs), which proves they reach the renderer in a drawable shape and that each declared reading changes the drawing. It proves nothing about how any of them looks. Worth one browser pass, and there is more to look at than usual: a portfolio now draws a kappa panel instead of an exceedance panel, a P&L carries break-even marks in both panels, a bivariate draws flat and no longer square, and the four controls (log, full range, return period, invert) are new words in a new place.

c) Fix what is shown where

The one that decides whether the rest was worth doing, and the least specified. Feeds back into (a): the switch cannot find its home until the tabs settle.

Reviewing this is what --library (a36) is for: point the server at a short .agg and the dropdown and hero gallery narrow to it, so a pass over the tabs is not a walk through 186 recipes.

uv run aggregate-api --port 8001 --library T:\tmp\short.agg

Still open from before and probably part of this: the Bounds tab has no content, and the standing REMINDER below about what is worth plotting.

plan-loss-lab-navigation — the answer to (c), and it absorbed the Bounds item above. Seven stages, one version bump each (a42 to a46): the Loss Lab rename, the capability payload that killed NA_TABS_BY_KIND and NA_MORE_BY_KIND, six nav groups with their own sub-tab rows, the Sharpen / PnL / Reset action row, the reinsurance entry box, Bounds, and the Narrative pane. (done; moved to dev/done/)

Left open by it, deliberately: the header's **Perspective** control. The
plan lists it under page-wide preferences and no stage owns it, so the
exhibit fetches take `insurer`, which is the plan's stated default. The
Tables switch beside it is live; this one is not built.

**Nobody had clicked any of it.** The gating was proved by
`check-nav.mjs` against real payloads and every route has tests, but the
six groups, the three derivation buttons, the cession box, the three
bounds forms and the Narrative pane had never been used in a browser.
The author's first pass found the Bounds group dark for every object
(fixed at a48, below), and the rest of that first pass became
`plan-ui-round-3`, immediately below.

plan-ui-round-4 — the second full pass (a55 to a61, with a54 the upstream catch-up that unblocked it). Eighteen items off dev/api-punchlist.md Round 4, six phases: the nav and its greying; the strip's verdict and history; the strip's facts; the action row; two hamburger preferences; the cession box and Quick Re.

Three of the six were the page telling the truth rather than new
features. The greyed group tab had been unable to draw its own tooltip
since a48, because Bootstrap's `.disabled` class carries
`pointer-events: none`. A clean gross under reinsurance tinted the strip
amber. A bivariate's status line said nothing but its name, because its
`bs` is two numbers. And `BuildResponse.warnings` had been declared from
the start and never filled.

**a60 was not in the plan.** Upstream removed the `plot_envelope` keyword
the api was passing, and rather than repair a route already marked for
deletion the author's call was to take matplotlib out: the envelope is a
chart document now, drawn by a new 2-D path in the adapter, and
`plotting.py` and the `/plot` route are gone. That is
`plan-plot-ir-api.md` item 1 and the bounds half of item 5, landed early.
*(done; move the plan to `dev/done/` once the browser pass below is made)*

**Nobody has looked at any of it.** Every phase is proved by tests,
`check-nav.mjs` and a node run of the adapter over real documents, and
not one of the six has been seen in a browser. Worth one pass: a
`grossceded` object for the tooltips, a bivariate for the status line, a
reinsured aggregate for the green verdict and Quick Re, and the Bounds
envelope for the new 2-D chart.

plan-ui-round-3 — the first full pass over the running site (a49 to a52). Roughly forty items off the author's punch list, taken after the first end-to-end run. Four phases: the shell and the status strip; the charts; the text and the tables; the reinsurance and pricing forms. (phases 1 to 4 done; the doc stays in dev/ until phase 5 lands)

**Phase 5 is not done and is blocked**, on
`aggregate_REFACTOR/dev/plan-loss-lab-round-3.md` phase A, which has not
started. Four items wait on it:

- `evaluate`, gross / net occurrence / net on the Evaluate form, which
  needs the library's basis keyword rather than more of the `_BasisView`
  shim.
- `alloc`, the per-unit allocation tables a reinsured portfolio silently
  loses, same keyword.
- ~~`cession`, pricing a cession with a distortion; the library has no
  glue between reinsurance and distortions at all.~~ **Done at a197**,
  `dev/done/plan-a197-layer-pricing.md`. The glue arrived in the meantime
  and then turned out not to be needed: `Aggregate.reins_price_df` prices
  gross / ceded / net and not the layers, so the quote distorts each
  layer's own annual distribution, reached by building the one-layer
  `ceded to` program `reins_program` renders. Quick Re now writes a
  `deposit` or a `rol` on every layer it cedes.
- ~~`replot`, the Reinsurance Plot leaf on a reinsured **portfolio**,
  which greys because `chart_reins` is registered for `Aggregate`
  alone.~~ **Answered, not fixed.** Library a244 rewrote `chart_reins`
  as the occurrence plot and states the restriction as the author's
  decision for 1.0: a book's units cede on different stages, so a
  portfolio-level gross / ceded / net triple would have to pretend they
  cede on the same one, and the aggregate cover is a separate contract
  with a separate picture. The leaf is dark there on purpose and
  `nav.js` says so.

Two things the round turned up that are the library's and are recorded
here rather than acted on:

- `plot_envelope(distortions='space')` is a documented value that does
  nothing: it matches neither of the overlay branches, so panels 2 and 3
  draw empty. a52 stopped asking for it and builds the list itself.
- `format_program(width=)` is accepted and ignored, which is the agg
  plan's `width` item.

SM

Bandwidth limited by caddy page with obscured url; special caddy install (done — rate_limit on the public /Q7M4Z9KP route, build-only, plus a friendly 429 card; see human-hints.)
Set up service to run on vps (done 2026-10-20: the systemd unit in human-hints, with AGGAPI_CACHE_MAX=100; refresh.sh restarts it)
Limit log2 <= 18 (AGGAPI_LOG2_CAP defaults to 18)
[ ]

Near term (get it healthy)

Backlog / ideas (unordered)

aggregate Loss Lab (aLL) work, staged

The plan behind these is the harness plan of 2026-07-29; one dev/plan-*.md and one version bump each.

plan-relink-library — catch up with aggregate a148 to a174 (→ 1.0.0a19). Examples read off build.recipes / Recipe instead of the deleted examples.agg; ?group=topic|kind|role + /examples/heroes; bivariate → bvagg; sev buildable; new /objects/{id}/meta; the actual_* / validation_description / recipes renames followed. (done; in dev/, move to dev/done/ when the stage set closes)
plan-all-branding — aggregate Loss Library (aLL) (→ 1.0.0a20). Two-line banner (name over "description to distribution" in grey small caps); name through title / meta / manifest / sw / About / Help / README / package description. No-dashes-as-punctuation rule adopted into CLAUDE.md, user-visible strings swept. (done; in dev/)
Dash tidy: the ~190 legacy -- glosses in internal docstrings and comments. Deliberately deferred out of a20 so it does not bury the staged diffs; the rule in CLAUDE.md says they are cleaned as their file is next edited. Promote to its own tidying commit if the author would rather have it in one go.
plan-echarts-exhibit — ECharts engine + two-panel Overview (→ 1.0.0a21). uPlot retired; web/src/charts/ with a served house theme (/v1/meta/style) and an exhibit per kind; density + EP panels with an exact index-aligned cursor link and dataZoom; new /objects/{id}/unit_density_df; sev gains a sampled density_df; Overview header block off /meta; duplicate Summary dropped from More. Bundle 187 → 353 kB gzip, now split four ways. (done; in dev/)
plan-exhibit-punchups — first inspection acted on (→ 1.0.0a22). Control row (log y / survival / full x / reference lines), sticky per browser; dataZoom rescales the y-axis; distortion and bivariate drawn square; program disclosure dropped; exhibit titles trimmed. Plus two bugs the inspection surfaced: a per-object read race that made the landing page 500 intermittently for portfolios, and xpnl-over-port / unresolved-reference turned from 500s into 422s. (done; in dev/)
plan-exhibit-punchups-2 — second pass (→ 1.0.0a23). Discrete densities drawn as steps (steps-mid); capital anchors are faint dashed verticals with top labels instead of points that could never sit on the curve; return periods rounded to integers. Smoke test reworked to replay captured fixtures so it needs no server. (done; in dev/)
Eyeball the Overview. Still outstanding. The smoke test proves the data reaches each chart in a drawable shape and that the panels hold the house aspect at both breakpoints, but not that any of them looks right, and the Chrome extension has not been connected in any session so far. Check all six kinds, both breakpoints, and the five toggles.
Watch the full-resolution render cost. (Closed by a37, and by an answer better than either option recorded here. sampling: 'minmax' reduces each series to the min and max per device pixel before the path is built, so a portfolio hands ECharts a few hundred points rather than six polylines of 65,536, and it re-expands on zoom because the sampler runs after the dataZoom filter. It was adopted to fix the step rendering, not the cost; the cost went with it. No server-side binning, as required.)
Draw atoms as atoms. Solved in a27 by dropping the binning entirely rather than by detecting atoms, which is not possible from the frame: a tall bucket and a point mass are the same number, so there is no threshold. The plots take every grid point; resolution='display' survives for the Density table. Gzip made the payload a non-issue (3.65 MB to 0.35 MB on the wire). (done)
The author is not sold on ECharts (raised a28, alongside the step / anchor / surface asks). Recorded so it is not lost: if the exhibits still do not earn their keep after a28's changes are seen, the question is whether the library is the problem or the chart designs are. Worth separating before any swap, because most of the a26 to a28 fixes were design errors (wrong rectangle measured, wrong default orientation, binned data) rather than anything ECharts did. a29 adds one more to that column: the reservation was aimed at the wrong rectangle for two kinds. Being settled by experiment: the a30 Plotly spike puts both engines on the same data behind a switch.
Plotly evaluation spike (→ 1.0.0a30). agg and port, behind a draw switch in the exhibit control row. The work turned out to be twoPanelData, the engine-neutral decision step both renderers now read, rather than the Plotly code. plotly.js-gl2d-dist-min and not basic: basic registers bar, pie and scatter only, and its SVG scatter would choke on the 2**16 points a27 ships. Parity asserted in the smoke test, including that Plotly log ranges are in exponents. (done; moved to dev/done/)
Look at the two engines side by side. This is the whole point of a30 and nobody has done it: the Chrome extension has never connected in any session, so no rendered chart from either library has been inspected. Build an agg and a port, flip draw, and decide. If Plotly wins, lift twoPanelData into a real IR with to-echarts / to-plotly adapters and shrink the bundle with a custom build registering only scatter and scattergl. If it loses, engine.js and plotly-panels.js come out and the dependency with them; keep the twoPanelData split either way. Update (a38): the real IR arrived, upstream rather than app side: the library's chart IR (dev/plan-chart-ir.md there) with the surface pilot at a38 (/chart/{name} route, chartdoc-to-echarts.js, surfaceGrid deleted). The two-panel family migrates onto it chart by chart as the library conversions land; twoPanelData remains the app side seam until then.
```
**Closed 2026-08-09.** The engine comparison is moot: the app assumes
IR-based charts always, so there is one renderer and no second engine to
weigh it against. `twoPanelData` is not lifted into an app-side IR, it is
deleted as each emitter lands. See `dev/plan-plot-ir-api.md`.
*(Done at a62: it and the whole of `exhibits.js` are gone.)*
```
greater_tables alongside the current tables (→ 1.0.0a31). One new GET frame/{which}.html route mirroring the .csv one, rendering server side on the real DataFrame so the row index sparsifies. renderExhibit stays and stays the default; GT is reached by ?tables=gt, no user facing switch. Row emphasis preserved as a <tr> class. (done; moved to dev/done/)
Compare the two static table renderers (→ 1.0.0a32). Settled by the author: "the GT tables are a delight". The comparison ended by replacing both sides, since greatest_tables (GT2) arrived with a semantic document IR and a client-side walker, which is a better answer than either. The hand-built renderExhibit and the 5.x html path are both gone.
Upstream greater_tables nit (closed by the same move). GT2 has no import side effects at all, by design, so the Pandas4Warning and the warnings-filter pollution are gone with the 5.x dependency.
Table view everywhere, one switch (→ 1.0.0a33). The header menu's Tables section steers the whole page; the Overview keeps a second affordance for the same value. Price frames travel with their documents, since a POST computes them. Gated on row count, so density stays interactive.
Look at the static tables. The walker has been verified end to end against real documents in Node (sparsified stubs, flags, negative cells, raw values in the CSV export), but nobody has seen one in the page. Build a portfolio, since the two level row index is the whole argument and an Aggregate does not show it. Then walk the tabs: More, Reins and Price all render statically now and none of those have been looked at either.
Math in tables and titles. The walker emits $...$ for page MathJax unless a katex object is passed, and the SPA loads neither, so a $...$ cell would render as literal delimiters. Nothing sends one today. It is the same decision as the chart labels, so make it once for both; it is now pressing on the chart side, because documents arrive carrying ChartDoc.tex (library a209) and the adapter drops it. Held with the chart half at dev/plan-plot-ir-api.md, item 8, where the ruling is that the IR emits a TeX and a plain-text form of every label and the renderer picks one.
Report to the greater-tables side: its requires-python = ">=3.13" forced this repo's floor up from 3.11. (Fixed upstream: 6.0.0a2 puts the floor back to 3.11 and says it verified 3.11, 3.12 and 3.13. Acting on it here is part of the catch-up below.)
Catch up with greater-tables 6.0. The sibling checkout ran from 1.9.0 to 6.0.0a4 while this repo was elsewhere, and a37 followed exactly one thing from it: TableSpec.formats became formatters, which had 13 tests red on main. Both harnesses (check-frames.py, check-adapter.py) are clean against 6.0.0a4, so nothing is broken; these are opportunities, not breakage. - Put requires-python back to >=3.11. a34 raised this repo's floor to 3.13 solely because GT declared it, and 6.0.0a2 reverses that. The classifiers go back too. - The deploy pin is now actionable. GT is publishing on the 6.0 pre-release line, so the standing note below about pinning 6.0.0rc1 has something to point at. Check what is actually on PyPI before pinning; pip still ignores pre-releases unless the pin is explicit. - 6.0.0a4 gives ratio_cols percent rendering again, and adds house format defaults (float_format, int_format, ratio_format, date_format, table_float_format). tables.py currently spells out its own per-column strings in FORMATS; some of that may now be a default worth inheriting rather than restating. - The walker moved with it. /v1/assets/gt-render.esm.js serves whatever the install carries, so 1.10's docs pass and 1.11's "string columns pass through as text" are already live in the browser and have not been looked at.
The 3-D surface is sugar, not substance (author's verdict after seeing it, a29). "Quite impressive. Fast. But more sugar than substance." It stays, since it is built and its chunk is lazy, so it costs the landing path nothing. Do not invest further there. The verdict sharpens the standing reminder above: the open question is what is worth plotting, not what can be rendered.
REMINDER for SM: better things to plot. The author's note, verbatim: "this was a general comment that i need to work out better things to plot. Just remind me if i don't come back to it." Not an api task. Raise it next time the exhibits come up.
Reinsurance-aware pricing. Landed in a27 as POST /v1/objects/{id}/reins_price. The a26 blocker (no public way to calibrate on the gross basis) dissolved once the author pointed at prob_loss_assets: a GridDistribution over the reins_density_df column supplies the (p, L, a) anchor, Pentagon.solve the premium target, and Aggregate.calibrate_distortions runs against a small view of that basis. No library internals reproduced. (done)
Upstream, small: let calibrate_distortions take its distribution. Landed upstream better than asked: calibrate_distortions(reins_view=...) at a223, fixed for the shapes this api needs at a250 ([Reins-Density-Fuzz]), and richer than _BasisView's three bases (gross, ceded, net, plus the occurrence intermediates on two stage programs, per Aggregate.reins_views). GridDistribution is public. The shim and _REINS_BASES are deletable on a sync to a250 or later, and their deletion is scheduled as phase A3 of dev/plan-pricing-exhibits.md. (done upstream; app-side deletion pending in that plan)
Hero gallery empty on a first page load (not reproduced). Reported by the author; populated on the next load. The server side is clean: /v1/examples/heroes answers 200 in ~2 s cold and returns all eight, including fired concurrently with /v1/examples and /v1/meta across four separate cold processes, and the service worker never touches /v1/*. a27 split the fetch and render error paths (they shared one silent .catch), added one retry after 750 ms, and logs on give-up. Next recurrence: check the browser console for [aLL] hero gallery.

Server policy (2026-07-29). The author starts and stops servers; the tooling here must not. dev/scripts/capture_fixtures.py and uv run pytest both drive the app in-process through TestClient, binding no port.

plan-examples-find — fuzzy find + hero sparklines (→ 1.0.0a24). Pinned search box in the dropdown and a Ctrl+K palette, both uFuzzy over name / kind / tags / note, flat ranked list while searching; hero cards upgrade from a placeholder gradient to a real density silhouette from a cached heroes/sparklines payload fetched after first paint. (done; in dev/)
Put the hero gallery in the Examples dropdown. a37 took the cards off the top of the page, where the author did not want them, and left the landing build: one role:hero entry is still picked at random and built, so the page arrives populated. What is gone is the showcase, and it should come back somewhere findable. pickRandom is still in main.js and /v1/examples/heroes/sparklines is still on the server for it; the SPA wrapper for the sparklines came out and is two lines to restore.
Replace the lede. web/index.html carries placeholder copy where the hero row was, sized at two lines. The author's to write.
Retire or repoint the hamburger's "Example source…" placeholder. (retired at a69. Shipped greyed-out in a17 for a source switcher that a159 made moot: the three shipped libraries are one. The live axis is ?group=topic|kind|role, reachable from the Examples menu, so there was nothing left to repoint it at.)
plan-more-tab — More as a tab, Bounds, Reins exhibit (→ 1.0.0a25). More is a pill whose pane carries a Reins-style sub-button row (Validation / Stats / Density / bs window / Info); Bounds ships greyed out; Reins gains the gross / ceded / net two-panel exhibit above its tables. Build errors now render on Overview rather than in a sub-menu. (done; in dev/)
Reins ceded density is a first-bucket spike on the shared linear window, since the ceded distribution is small next to the gross. The log y and full x toggles cover it for now; if it stays hard to read, give each series its own crop rather than one shared window.
Bivariate 3D viewer (→ 1.0.0a28). echarts-gl surface on the joint density behind a dynamic import (own lazy chunk, 165.9 kB gzipped), linear and log height, falling back to the heatmap without WebGL. The downsample to 128x128 is client-side after all: the joint payload is already only 2,560 rows, so a server-side reduction would have saved nothing and added a route. (done)
Nobody has seen the 3-D surface render. surfaceOption is pure and the smoke test checks the mesh is complete and has relief, but WebGL is not exercisable offline. First look should check: does the ridge read on CopulaWindFlood, is the log floor the right depth, and do the wall projections show the marginals usefully or just add clutter.
The percentile reference lines come off: dev/done/plan-no-reference-lines.md, canonical here and pointed at from the library's dev/done/. Executed both sides the day it was written, 2026-08-13: LIB a271 deleted CAPITAL_ANCHOR and LEE_ANCHORS and the marks they built, app a98 is one tooltip and two stale comments, since the app draws what it is served. What is left is the mean, and break even on a P&L. Browser pass done on all three kinds. One thing to know: the mean's label changes side, taking the right as the only vertical under the rightmost rule, and the rule was left alone. (done)

The reflected reading, verification only: dev/done/plan-chart-reflect.md, the app half of the library's plan of the same name, paired rather than symlinked since each repo owns its side. Both halves are written: the app at a96, the library at a269, which is in the library's working tree rather than committed. The upstream question closed the app's way, a269 composing the two probability readings with one flip and no REFLECTED_RETURN_PERIOD_MAP, so reading-map.js and its tests need no edit.

Verification steps 1 to 3 ran clean on the a269 sync, 2026-08-13:
`dev/fixtures/charts.json` re-captured and `smoke-charts.mjs` all clear,
with `reflect` offered on exactly the six documents that declare a
complement axis (`agg`, `severity`, `distortion`, `pnl`, `reins`,
`envelope`) and absent from `port` and the joint surface.

**What is left is the browser pass**, step 4, whose load-bearing check is
that `reflect` with `return period` opens out the *other* end of the
curve. If it changes nothing there, a double flip is back somewhere.

The joint surface, end to end: dev/done/plan-3d-plot.md, three parties, the canonical copy here and a symlink in the library, both moved to dev/done/ 2026-08-21. This side is complete: a72 the api half (section 3, the three query parameters, the ceiling setting and the chart document cache), a73 the foundation (section 4.1's one decode, section 8.3's node --test runner, the echarts-gl pin), a74 the drawing geometry, a75 the two halves read against each other, a76 the relief and both author rulings, a77 the look, a78 the controls, a79 the cuts, the marks and the walk, a80 the window box, with later fixes through a92. Section 4.2.1 was ruled 2026-08-12: keep the full grid rule, so the derived quantities stay this side and the library emits the whole reduced grid with window as the drawing range inside it. Section 4.2.1.1, the library serving the total lattice and kappa, is declined by the same ruling. What is left is entirely library, two edits carried by aggregate_REFACTOR/dev/plan-3d-plot-LIB.md: edge = "mid" and the whole reduced grid. Nothing here changes when they land, verified 2026-08-21 against windowRange and the surface-grid.js decode. The one chore on landing is re-capturing dev/fixtures/charts.json and re-running smoke-charts.mjs. Until then every conditional and every mark is computed on the part of the grid that is on screen, up to 4.9 percent out on kappa, and the window box set to 0 is the escape hatch. The prototype stays in dev/prototypes/joint-surface/ under section 8.0's rule, as the reference the shipped surface is diffed against.
The Pricing pane through the official channels: dev/plan-pricing-exhibits.md (canonical here, symlinked in the library), FINAL 2026-08-12. Calibrate, Allocate, Evaluate subtabs served by pricing.calibrate / pricing.allocate / pricing.evaluate exhibits keyed on new result objects; a debounced pentagon preview line; [Calibrate] replacing [Price]. The library's five phases landed at aggregate 1.0.0a259 to a263, with execution notes and nine recorded divergences in the library's dev/plan-pricing-exhibits-LIB.md. Three phases here, all landed: A1 routes a83, A2 pane a84, A3 deletions a85. pricing.py is thin runners with no pandas, and tables.FORMATS holds the two sharpen entries alone, which the sharpen leaf's own cutover (above) deletes.
SpaceMouse on the joint surface, and the surface out of the app: dev/done/plan-spacemouse.md, app repo only, CLOSED 2026-08-12. Four phases: a86 the mesh export, a87 the probe page, a88 the device layer plus the camera integrator, a89 the feel, then a90 the probe findings from the author's unit (the combined 12-byte layout, travel of plus or minus 350, one multi axis collection, no double action from 3DxWare) with the two bugs that run found, a button mask that could fire its own camera reset and a vertical pan running against the hand, and a91 the color ruling, GLB with per vertex viridis read against the colorbar. The relief's row carries mesh | .glb | .obj | .stl, spacemouse and feel; Chromium only, feature-detected, greyed elsewhere, iPad path untouched and the echarts-gl pin unmoved. Five residual readings (session lifecycle under the puck, sleep and wake, zoom and twist direction, the GLB in 3Dconnexion's viewer and the STL in a slicer, the greyed control in Firefox, Safari and iPad) are recorded in the closed plan's "What stays unread" section, and come back as punchlist entries if they matter.
The app on an iPhone and an iPad: dev/done/plan-idevice-ui.md, FINAL 2026-08-14, extended and re-verified 2026-08-21. Five defects, SPA only, no api or library change. All three code phases landed 2026-08-21. a115: the group and leaf strips are laid out at their selected width so they stop moving under a thumb, and touch-action: manipulation on body stops a double tap zooming the page. a116: a vertical up / readout / down control in the editor's right margin, the counter reversed to count up JupyterLab style, the cap from 20 to 500 with a 256KB ceiling, the unbuilt draft stashed and handed back, the walk ending on a document change rather than on any keydown, Ctrl bound beside Mod and the key hints naming whichever modifier the device reports. a117: charts/touch.js, a capture-phase shim giving touch events the offsetX that echarts-gl reads off the raw event, so a tap on the 3-D surface picks again, and one resetView reached from the button, the double click and the puck, with the reset button now on every chart and last in its group.
The iPad round for the above, phase 4, owed. Nothing in a115 to a117 can be verified from a Windows desktop: the pick defect lives in a branch Safari takes and desktop Chrome does not, and the modifier finding rests on a platform test that is false on Windows. Fifteen minutes against the six checks in the closed plan's section 6. Two things it may send back. The tap in check 5 can fail in its second half, a deliberate tap over 300ms being refused by zrender no matter how good the coordinates are; the recognizer for that is sketched in the plan and deliberately unbuilt. And section 6.1 is unconfirmed: whether iOS Safari applies focus zoom to a contenteditable as it does to a form control, which if it does means every tap into the editor on an iPhone zooms the page, a larger source of the "floating around" report than the double tap and one that touch-action does nothing about. Its fix is pre-specified, a 16px .cm-content under 576px in the editor theme, and takes its own bump. Also open for a ruling, from the a117 notes: resetView restores the relief's readings on a surface only, so contours toggled off on a heatmap survives a reset.
The 2-D control strip goes per panel, and three axes stop lying (→ 1.0.0a121, against aggregate 1.0.0a314). dev/done/plan-2d-punchups.md, both halves executed 2026-08-21. log split into log x and log y, the seven readings moved from one centered row into one group per panel keyed on the panel id, full range stopped falling back to the drawn extent so a probability axis stops being labeled to 1.5, and the return-period window stopped being discarded outright, which was four independent faults on two lines of xyPanel. Properly fixes the a-generation punch item "log y on rh plot triggers reshape/draw of left plot" (dev/api-punchlist.md:238). Two further rulings taken on execution and recorded as decisions 10 and 11: the log release acts on a density ordinate alone, keyed on the axis unit so it survives invert, and a panel offers on every axis that can reach a position rather than on the one occupying it now. No VIEW_KEY bump. Still owed: the browser pass, sections 5 and 6 of the plan's verification, which is the group alignment at both sides of WIDE_PX and the sticky state carrying across documents.
The companion axis auto-fits over data the return-period ladder no longer shows. Raised in review 2026-08-21, not ruled, recorded rather than built. Under a period reading the companion is given no window so echarts fits it, which was right when the period axis showed everything and is loose now that aggregate 1.0.0a314 ends the suggested ladder at 1-in-10,000: the visible part of the curve fills 41% to 100% of the axis across the fixtures and 47% on the author's own program. The fix is to take the companion's extent over the points whose period coordinate falls inside the period window, about three lines in xyPanel. See the "left behind" section of dev/done/plan-2d-punchups.md.
A linear return-period axis is labeled from zero. niceWindow rounds the declared (1, 10000) outward to a tick interval and floors the low end at 0, so the first tick reads 1-in-0. New at a121 and harmless: the axis was unconditionally log before aggregate 1.0.0a314, so the question could not arise. Not fixed with it because clamping niceWindow to a declared minimum would act on every axis in every document.
Should the per-panel groups carry their panel title in the wide layout too? a121 labels them only when stacked below WIDE_PX, on the grounds that panel titles are long ("Probability mass function") and the panel headings are already on screen right above them. Turning them on everywhere is one condition in renderControls. A styling call, open.
The Examples dropdown reads the library in file order: dev/done/plan-examples-dropdown.md, complete across a122 to a124, against the LIB half at aggregate 1.0.0a320 (Recipe.seq, Recipe.as_read). A122: the payload is one flat file-ordered list with pills and facets, the app's three ordering tables are deleted, and the editor receives the entry's own text rather than a canonical re-render. A123: the three-line row, the pill filter bar with counts, localStorage persistence shared with the Ctrl+K palette, and the Ctrl+Shift arrow ring walking the filtered list. A124: out-of-order search, measured on the real payload, so the fzf port the plan held in reserve is not needed. Owed upstream, ruled 2026-08-24: topic:pnl and topic:distortion duplicate their entry's own kind on fourteen rows, so the pill row spells the same word twice; the author is removing the two tags from library.agg, and nothing app side changes when they go.
Catch the app up to aggregate 1.0.0a333: dev/done/plan-library-skew-a333.md, one bump at a140. Turned up by the [Sync] step of the approximation tab plan, which is why it exists as its own bump: nine backend tests were failing on library skew and none of them belonged to that feature. One real defect, a bivariate's moments coming back empty since library a330 restructured stats_df, and eight test expectations describing surfaces the library has restated (the two sided return period ladder, the joint surface's centered coordinates and whole placed mass). Nothing owed upstream: two of those tests were reading detail and the emitted mass against quantities the library never promised, and both now assert what _joint_surface documents.
The Approximation tab: dev/done/plan-approximation-tab.md, one bump at a141. The More row's new leaf, between Tail behavior and Window, with a Plot / Summary switch inside the pane rather than two sibling pills: the navigation is two levels and check-nav.mjs guards that shape, and the two are readings of one question so they keep one address. Consumes LIB a331 to a333 and needed no server change at all. Eight divergences in the plan's execution notes, the load bearing ones being a header hook on loadExhibitLeaf (the switch has to be rebuilt inside draw, which re-runs on both table flips) and the first chart instance pane-more has ever held, which disposePaneChart and clearPanes had no branch for. The Re rename rode in the same commit on the author's request. Open, upstream and not blocking: the chart payload is 16.3 MB against the agg chart's 2.3 MB, thirteen series over the full 65,536 point grid, taken to the LIB side as a sizing question. Nothing app side is owed either way, since a smaller emission lands here with no change.
The Pr Ruin pill: dev/done/plan-pk-tab.md, one bump at a142, consuming LIB a339 to a341 (the engine lift, eventual_ruin with RuinResult and the ruin exhibit, the ruin chart with its downsampling), which landed concurrently from the same plan; the LIB execution notes were read before any app code. Last pill in the Pricing row, after Evaluate: the shared pricing form plus a probability of default box, over the two-panel chart and the exhibit stats strip, live to the form through one debounced POST. Divergences in dev/done/plan-pk-tab-API.md, the load bearing ones being no can_ruin flag (the pill gates on the chart, the Plot mechanics), the new POST route the plan's "nothing new" line underestimated, and sample: true as the wire spelling of the seed re-roll. Nothing owed upstream.
The reinsurance structure diagram: dev/done/plan-structure-chart.md, three bumps, consuming LIB a349 to a351 (the 'tower' panel kind and TowerBlock on the chart IR, the structure emitter, the matplotlib renderer). The library half landed first and deliberately left the SPA drawing nothing at all, since it pinned CHART_IR_VERSION = 2 and refuses anything later. a152 [Tower-Panel] closes that window and draws a tower, with the one row ratio layout panelLayout had no notion of; a153 [Diagram-Leaf] puts it at the head of the Re row and lights that group for a reinsured P&L, which is the first time that group has been live for one; a154 [Structure-Lee] carries lee and annotate on the chart route, without which the Lee curves could not be asked for at all. Three rulings and nine divergences in the plan's execution log, the load bearing ones being the route parameters (the plan called the app side "the translator plus one nav leaf" and there was no way to reach half of what it described) and the layout, which is in the adapter and not in mount.js where the plan looked for it. Nothing owed upstream. Eyeballed at a154 and three defects came back, all of them the library's loss window; they were specified in aggregate's dev/done/plan-tower-window.md, landed there at a352 and a353, and consumed here at a155 [Tower-Log] along with the app's own half of the log reading: one decade floor per quantity axis rather than per panel, the floor strictly under the smallest value drawn, bands and boundary ticks that cope with zero, log y offered once per axis and acting on every panel naming it, and a torn top edge on an unlimited block. Eyeballed on log and approved, with one defect: the switch appeared to drive one stage's quantile curve and not the other's, which turned out to be a reading stored against a panel that no longer offers one, from the window where a154 ran against a352 documents. Fixed at a156 [Stale-Companion-Reading], along with labeling the two control groups by their axis, since side by side they were identical and unlabeled.

Raised with `aggregate` (not fixed here)

Raised with `csv-grid` (not fixed here)

The status line is written before maxRows is applied. 3.10.1's _install runs refresh() (body and status) and then _applyHeight(), which re-renders the body windowed but not the status, so a grid that windows keeps "showing rows 1–2,048" in its status bar. loadAllRows calls renderStatus() once the load promise settles. Ask: _applyHeight refreshes the status after it re-renders. Raised 2026-10-17, 1.22.0.
No way to default the export to raw values. CsvGrid 3.9.0 builds the "Formatted values" checkbox with checked = true hard coded (csv-grid.es.js:824) and _runExport reads it live, so there is no constructor option and no export default to set. mountGrid clears both boxes (Copy and Save) by reaching into the widget's DOM after construction, which works only because _buildScaffold builds them in the constructor. Ask: an exportValues: 'raw' | 'formatted' option. The workaround deletes the day it lands. Raised 2026-08-10, a64.

Navigation Menu

Repository navigation

TODO.md

TODO.md

TODO / Roadmap

Landed

Next up: three things, and then we are in good shape

a) One payload per table, and move the switch — done (a34, a35)

Exhibits endpoint, remaining slices (a39 landed the routes)

b) Consolidate the graphs on ECharts, and fix the options

c) Fix what is shown where

SM

Near term (get it healthy)

Backlog / ideas (unordered)

aggregate Loss Lab (aLL) work, staged

Raised with `aggregate` (not fixed here)

Raised with `csv-grid` (not fixed here)

FilesExpand file tree

TODO.md

Latest commit

History

TODO.md

File metadata and controls

TODO / Roadmap

Landed

Next up: three things, and then we are in good shape

a) One payload per table, and move the switch — done (a34, a35)

Exhibits endpoint, remaining slices (a39 landed the routes)

b) Consolidate the graphs on ECharts, and fix the options

c) Fix what is shown where

SM

Near term (get it healthy)

Backlog / ideas (unordered)

aggregate Loss Lab (aLL) work, staged

Raised with aggregate (not fixed here)

Raised with csv-grid (not fixed here)

Raised with `aggregate` (not fixed here)

Raised with `csv-grid` (not fixed here)