[MarkusNeusinger]

Summary

This PR audits how faithfully the agent chat keeps pasted data and fixes every silent precision loss it found. It follows the owner rule of 2026-10-15: what the user pastes is what the data holds, or a warning or change line says what changed and how often. Silent precision loss is treated like silent corruption.

Audit

Path	What can be lost	Before	After
code/loader.py: run-form pd.read_csv (exported plot.py, sandbox render)	float digits	No float_precision. On pandas 3.0, about 23 % of realistic 17-digit floats were read 1 ulp off, so the plot did not draw the values in data.csv.	float_precision="round_trip" whenever a number column exists. Test: 2,000 random floats of every magnitude plus extremes read back bit-exactly.
parse.py _as_numbers: CSV number cells	digits past float64 (0.1234567890123456789, 9007199254740993 beside decimals, subnormals, 1e-400)	Silently rounded.	Stored as the nearest float (repr), with a warning that counts the cells: "N values with more significant digits than a number holds (15 to 17) are stored rounded".
parse.py _decode_literal: JSON and Python-literal floats	digits	json.loads and literal_eval rounded them before typing, with no warning.	The written text is kept (_SourceFloat), so the same count applies. JSON 1e400 is now text instead of vanishing as missing.
parse.py integers past int64	last digits of IDs and counts	Became float (1.2345678901234568e+22).	The column stays text with a warning. set_type number can still convert it, and its change line counts the rounding.
parse.py profile min/max of integer columns	digits past 2^53	float	Exact int (ColumnProfile.min/max: int | float | Cell).
parse.py trailing zeros, +, exponent, -0	spelling	1.50 becomes 1.5, 1e3 becomes 1000.0, -0 becomes 0	Unchanged by design: the value is identical (shortest round-trip form). See the findings below.
parse.py thousands and decimal-comma normalization, % and currency	value	Already warned (notes and unit-moved warnings).	Unchanged, now covered by an exact-digits test.
parse.py text cells: surrounding whitespace	characters	Silently stripped.	In text columns, one warning counts the cells. Space-aligned tables (layout padding) and number or date columns get no warning.
parse.py text cells: None, -, null, …	the whole cell (a possible category)	Silently became missing.	In text columns, one warning counts the cells and names the tokens. Number and date columns are unchanged (the usual missing markers there).
parse.py NFC	code points	Silent.	One warning counts the changed lines.
parse.py zero-width and bidi characters, header canonicalization, 200-character cell cap		Already warned or refused.	Unchanged.
parse.py thinning of list fields over 20,000 rows	rows	Already warned with counts.	Unchanged.
prepare.py set_type number	whole numbers past 2^53, many-digit decimals	Read through float (9007199254740993 became …992), silently.	Whole numbers are read exactly through Decimal (Int64). The change line counts rounded cells.
prepare.py aggregate sum of Int64	the whole value	pandas wrapped around silently (2 × 9.2e18 gave -1616).	Summed in Python ints. Past int64 the result is float, and the change line counts the rounded values.
prepare.py derive cumsum of Int64	digits past 2^53	Computed in float64, silently.	Computed in Python ints and stays whole. Past int64 the change line counts the rounded values.
prepare.py derive sum and difference past int64	digits	Silent fallback to float.	The change line counts the rounded values.
prepare.py pivot_longer of Int64 beside float64	digits past 2^53	Silent cast.	The change line counts the rounded whole numbers.
prepare.py derive ratio, share, percent	digits	Full float, written with repr.	Unchanged, now covered by a test.
prepare.py _cell_text: date-times	sub-seconds	timespec="seconds" dropped them silently.	moment_text: microseconds or nanoseconds when present. The offset is kept.
prepare.py set_type datetime without a format	month and day order	2024-06-01T13:00:00Z was read as 6 January (dayfirst=True).	Year-first columns are read year, month, day.
prepare.py set_type datetime with mixed UTC offsets		Uncaught ValueError, so the tool crashed.	type_mismatch refusal. The times are not converted.
prepare.py _read and _write, render/harness.py frame_csv		_read already used round_trip. to_csv writes repr (verified bit-exact on 20,000 random doubles).	Unchanged.
parse.py dates, _iso, loader parse_dates	seconds and range	Seconds are kept. Years 1 to 9999 read as M8[us] on pandas 3.	Unchanged. Fractional seconds and offsets at parse time belong to the sibling PR (feat/agent-file-drop-binary).
Page preview (DataPanel PreviewTable)	display	Shows the canonical cells verbatim, with no number formatting. Cells over 40 characters are cut with …, and whitespace runs are collapsed for display.	Unchanged. See the findings below.

Owner decisions (2026-10-15), applied

1. Trailing zeros and spelling. The value is enough (1.50 becomes 1.5), but one warning now counts the numbers whose trailing zeros, exponent or plus sign went, and names their columns. Spelling the canonical form only adds (2 as 2.0, .5 as 0.5, 1e-7 as 1e-07) is not counted.
2. Different UTC offsets are converted to UTC with a change line that names the offsets, consistent with the parser since feat(agents): read dropped Excel, ODS, Parquet, Feather, Arrow and NumPy files #12194. This now also applies to set_type datetime. Times with and without an offset in one column are refused.
3. Preview. The server keeps each preview cell's whitespace (parse.preview_cell). The data panel draws a line break as ↵ and each space of a run as ·, muted and with screen-reader text. The model-facing sample and top values stay one line.
4. Profile mean and median at six significant digits are kept, for the model only; this is documented.
5. Integers past 2^53 in the browser are kept as is, documented. This only affects the binding hint.
6. The CSV-delimiter space exemption stays: there is no warning for 1, alpha.

Merge with #12194

origin/main is merged in. #12194's _cell_text (value.isoformat(), which writes every fraction digit and the offset) replaces this PR's moment_text, and both test classes are kept. The stress case duplicate-headers now expects 4 warnings, including the spelling line; it was regenerated with make_stress. The parse-only fixture test ignores the spelling line.

Frontend

The change is app/src/sections/agent-chat/previewCell.ts plus its use in DataPanel, with Vitest coverage in previewCell.test.ts and DataPanel.test.tsx. App gates on Node 24 pass: lint, Prettier, type-check, 1,114 tests and the build. I did not drive the page in a browser in this session.

Test plan

• TestFidelity in test_parse.py with generalized fixtures: random decimals of 1, 7 and 15 digits in CSV, JSON, written JSON and Python literals keep their exact value with no warning; 19-digit decimals plus edge cases are counted exactly; int64 limits are exact and larger integers become text; NFC, whitespace and missing-token warnings; no warning for aligned padding or number columns.
• TestFidelity in test_prepare.py: 500 random floats of every magnitude plus integers past 2^53 survive operations exactly; set_type, aggregate sum, cumsum and sum overflow, the pivot_longer count, sub-seconds and offsets, the year-first fix and the mixed-offset refusal.
• test_loader.py: bit-exact read-back of 2,000 random floats plus extremes through the run form.
• uv run pytest tests/unit -n 8: 11,318 passed. ruff check and ruff format --check on the repo, mypy api core agents and the changelog check all pass.
• No app/ change, so no frontend gate applies.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GdMSqLR5ww4ji74EUSmk9M

[codecov]

Codecov Report

❌ Patch coverage is 96.96133% with 11 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
agents/anyplot/data/parse.py	94.55%	8 Missing ⚠️
agents/anyplot/data/prepare.py	98.42%	3 Missing ⚠️

📢 Thoughts on this report? Let us know!

[Copilot]

🟡 Changes recommended

Parsing performance, incorrect running-total grouping, and incomplete loss warnings remain unresolved.

5 open findings

• Avoid quadratic AST source scanning during paste parsing · New
• Preserve numeric grouping semantics for signed zero · New
• Preserve digit fidelity when parsing underscored floats · New
• Preserve leading-space evidence in CSV text columns · New
• Ensure rounding notices survive truncated operation messages · New

What changed in this PR

Improves data fidelity across agent-chat parsing, preparation, and exported plots by preserving precision or reporting changes.

Changes:

• Adds exact float loading and counted normalization warnings.
• Preserves integer calculations and timestamp precision.
• Expands regression coverage and documents fidelity guarantees.

File	Description
tests/​unit/​agents/​evals/​test_stress.py	Checks round-trip loader generation.
tests/​unit/​agents/​data/​test_prepare.py	Tests operation precision and timestamps.
tests/​unit/​agents/​data/​test_parse.py	Tests numeric fidelity and normalization warnings.
tests/​unit/​agents/​code/​test_loader.py	Tests exact float read-back.
docs/​concepts/​agent-network.md	Documents fidelity behavior.
changelog.d/​agent-data-fidelity.md	Records user-facing fixes.
agents/​evals/​stress.py	Aligns evaluation loading with production.
agents/​anyplot/​schemas.py	Allows exact integer profile bounds.
agents/​anyplot/​data/​prepare.py	Preserves calculation precision and reports rounding.
agents/​anyplot/​data/​parse.py	Retains numeric source text and counts changes.
agents/​anyplot/​code/​loader.py	Enables round-trip float parsing.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🟡 Changes recommended

Oversized exponents can cause unhandled conversion errors, and mixed date inputs can still silently change valid timestamps.

2 open findings

• Handle oversized exponents with a structured PrepareError · New
• Prevent tolerated invalid dates from changing column date order · New

5 resolved since last review

• Preserve numeric grouping semantics for signed zero
• Avoid quadratic AST source scanning during paste parsing
• Ensure rounding notices survive truncated operation messages
• Preserve leading-space evidence in CSV text columns
• Preserve digit fidelity when parsing underscored floats

🧠 Review effort: Balanced