Table lens: derived expr columns + author sort, computed at read time

[snudurupati]

Why

Some column values are derived facts that decay with time — the motivating case is a lead's freshness score in the crm Daily-leads table. Freshness is a function of a stored post date and now; storing the score in the graph means rewriting every row every day (1000 leads → 1000 daily touches). The right shape: the graph stores the durable fact (the date), and the notebook derives the current value on every render.

What

column.expr — an optional per-column expression evaluated against each row at query-refresh time; the result is injected under the column's key (so numeric right-align etc. just work). column.precision rounds it.

The expression language is deliberately tiny and safe — arithmetic, parens, and three builtins; no ambient environment, no property access, not eval:

• num("col") — row field as a number
• days_since("col") — fractional days from a date/Z-less-datetime field to now
• tier(x, t1,v1, t2,v2, ..., default) — step function for decay tiers

Worked example (current lead rank from a stored post date + stored component scores):

- key: current_rank
  label: Score
  precision: 2
  expr: >
    0.35 * tier(days_since("s.observed_at"), 1,1.0, 3,0.75, 7,0.55, 14,0.35, 30,0.2, 0.05)
    + 0.25 * num("eq.summary") + 0.40 * num("ea.summary")

table.sort — {key, dir} pre-render sort, needed when display order depends on a derived column the source query cannot order by. Numeric-aware, stable, nulls last; string comparison is pinned to a fixed locale (Intl.Collator("en")) so order is identical in the Node TUI and any browser.

Where

Evaluation happens once in core's buildRuntimeProps (not in the renderers), so web and TUI show identical values and the components stay presentation-only. Failure mode: a bad expr, unknown function, or missing field yields null (blank cell) — never NaN, never a render-time throw.

Tests

New expr.test.ts covers precedence, the builtins, decay-tier selection, null-on-garbage (including window.alert(1) rejected), derived+sort integration, and no-op passthrough when unauthored. Full workspace: 5 packages build, all suites pass (core 131, web 13 files).

🤖 Generated with Claude Code

Greptile Summary

This PR adds read-time derived columns and author sorting for table lenses.

• Adds a small expression evaluator for table column formulas.
• Injects derived values into table rows during core runtime prop building.
• Adds precision handling for derived numeric values.
• Adds stable, nulls-last table sorting with fixed-locale string comparison.
• Adds tests for expressions, invalid inputs, derivation, and sorting.

Confidence Score: 5/5

This looks safe to merge.

• No blocking issues found in the changed code.

Important Files Changed

Filename	Overview
packages/core/src/catalog/expr.ts	Adds the expression parser and builtins, with invalid numeric results returning null.
packages/core/src/catalog/lenses/table.ts	Adds table expr, precision, and sort props, then materializes derived rows with stable sorting.
packages/core/src/catalog/index.ts	Runs table derivation and sorting while building Table runtime props.
packages/core/src/catalog/expr.test.ts	Adds tests for expression evaluation, bad inputs, derived values, and table sorting.

_{Reviews (4): Last reviewed commit: "expr: tier() rejects non-finite argument..." | Re-trigger Greptile}

Context used:

• Context used - CLAUDE.md (source)

[snudurupati]

Addressed the boolean-coercion review comment in bccc7fe: num() now only passes numbers through and coerces non-empty strings; booleans and whitespace-only strings evaluate to null like other bad input. Added test coverage for both cases.

[greptile-apps]

Want your agent to iterate on Greptile's feedback? Try greploops.