FREEZE / REPORT constants first, --freeze, then replay: d30_venus_2026_11_04.
Contract rotated with previous = audit_wave2_2026_11_04 (kept byte-for-byte);
freshness test and frontend confirmation-gate test updated.
Versus audit_wave2_2026_11_04 (product accepted these costs 2026-11-04, no
tuning):
- sealed 20: 17 identical / 3 changed (case 5 rank 1 -> 3, case 18 12 -> 10,
case 14 reasons only); top_1 0.50 -> 0.45, MAE 6.65 -> 7.05; top_3 0.50 and
coverage 0 unchanged. All from the D30 table.
- reported offset 900: 650 identical / 250 changed; top-1 hits at
+-15/30/60 14/17/15 -> 4/3/0 of 300 (all from Venus roles); top-3 and
delivery coverage unchanged at every radius; mean minute error
19.79 -> 19.51 (from D30).
Attribution from single-fix replays on git-archive copies: D30 alone 3 sealed
/ 166 offset trials, Venus alone 0 / 108; they overlap only in case 20.
BUG-1560 and BUG-1570 resolved; ledger and CHANGELOG lines added.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
G (BUG-1560/1561: varga.py, divisional_charts_extended.py), H (BUG-1566/1567:
jaimini.py) and J (BUG-1270/1576: shadbala.py) each changed frozen identity
files and each froze its own audit_g / audit_h / audit_j record, which only
matches that branch's tree. Per ERR-110 the merged tree is frozen once as
audit_wave2_2026_11_04 (FREEZE/REPORT constants first, --freeze, then replay).
The contract now points at audit_wave2 with moon_paksha_2026_11_04 as the
previous record; the per-branch records are kept byte-for-byte and named as
intermediate in the superseded_reason notes. Freshness test and frontend gate
test updated.
Compared field by field with moon_paksha_2026_11_04: sealed holdout 20/20
trials identical, metrics unchanged; reported offset 900/900 identical,
summary unchanged; changed_trial_count vs 2026-09-20 stays 8 / 581.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
Freeze contract, FREEZE/REPORT constants and gate tests keep the G state
here; the merged tree is re-frozen once as audit_wave2_2026_11_04 next.
J's audit_j records are kept byte-for-byte as history.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
Freeze contract, FREEZE/REPORT constants and gate tests keep the G state
here; the merged tree is re-frozen once as audit_wave2_2026_11_04 at the end.
H's audit_h records are kept byte-for-byte as history.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
BUG-1560 / BUG-1561 changed scripts/varga.py and
scripts/divisional_charts_extended.py, both in the frozen scoring identity.
Per ERR-110 new records were created with each script's --freeze and
replayed as audit_g_2026_11_04; scripts, contract, freshness test and the
frontend gate test point at them. Prior records are kept byte-for-byte.
Compared field by field with moon_paksha_2026_11_04: sealed holdout 20/20
trials identical, metrics unchanged; reported offset 900/900 identical,
summary unchanged; changed_trial_count vs 2026-09-20 stays 8 / 581.
(The D16/D20/D27/D45 and extended-module charts are not read by scoring.)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
Data audit wave 2 (G): 19 fixed charts, Raman and Lahiri, mean nodes, both
engines set explicitly. Against PyJHora method 1 ("traditional Parasara") on
the same longitudes, varga.py matched D16/D20 26%, D27 25%, D45 14%, and
divisional_charts_extended (chart page, consultation varga spectrum) matched
D16/D20 7%, D27 14%, D40/D45 25%, D30 50%. jyotishganit agrees with PyJHora
190/190 on every one of these, and both follow BPHS ch. 6.
- varga.varga_map: D16/D45 movable/fixed/dual from Aries/Leo/Sagittarius,
D20 from Aries/Sagittarius/Leo, D27 fiery/earthy/airy/watery from
Aries/Cancer/Libra/Capricorn (were odd/even starts). BUG-1560.
- divisional_charts_extended: D16/D20/D27/D40/D45 signs now read varga_map;
D30 uses the BPHS signs (odd 25-30 Libra; even Taurus, Virgo, Pisces,
Capricorn, Scorpio). BUG-1561.
- trimshamsa_d30: even-sign table copied from jyotishganit swapped Jupiter
and Saturn; now 12-20 Pisces, 20-25 Capricorn, 25-30 Scorpio. BUG-1562.
- shodasavarga guide: D4/D10/D24/D27/D30/D40 rule text corrected.
Held back: varga.py even-sign D30 has the same defect, but it feeds
rectification scoring. Fixing it moved the sealed holdout (top-1 0.50 to
0.45) and 166/900 offset trials and turned the memoization goldens red,
which per BUG-1260/1275 means a scoring-13 identity: a product decision.
It carries a KNOWN DEFECT comment and a strict xfail test.
After: same-longitude agreement with PyJHora is 190/190 for D3-D60 in
divisional_charts_extended and for every varga.py chart except D30.
D9/D10/D12, karakas and Arudha are unchanged; the evidence card check
(27 cards, 765 fields) shows 0 differences before and after.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
Dig Bala took |powerless point - planet| / 3 without folding arcs above
180 degrees, so 35 of 140 values on the 19 benchmark charts plus Steve Jobs
exceeded the 60 Virupa maximum (up to 98.98). BPHS 27.7 ("hitva
shadbhadhike sati"; Santhanam: "If the sum is above 180 degrees, deduct the
sum from 360") and Raman, Graha and Bhava Balas section 44-45, Ex. 14-15
(Kuja arc 167 05, Digbala 55.70) both fold. PyJHora's default _dig_bala has
the same unfolded formula, which is why the comparison always "matched";
its own _dig_bala_another, VedAstro source, jyotishganit and saravali fold.
After the fix all 140 values equal PyJHora _dig_bala_another.
Drik Bala summed the aspecting planets in set order. Each Sphuta value is
already rounded to 0.01, so the quarter often sits on a .xx5 edge and the
result moved by 0.01 with PYTHONHASHSEED. Sum whole hundredths in sorted
order and round the quarter half away from zero; 14 of 140 values move by
0.01. Formulas, tables and benefic/malefic classification are unchanged.
New tests lock Raman's Standard Horoscope Ex. 13 (Sthana), Ex. 15 (Dig)
and Ex. 55 (Drik), the fold bounds, and seed-independent Drik. Updated
expectations: test_shadbala_precise_context (PyJHora default Dig carried
the bug) and the Jobs Moon total in test_shadbala_moon_paksha_convention.
shadbala.py is in the rectification frozen identity, so per ERR-110 both
fixes were frozen and replayed once as audit_j_2026_11_04. Compared field
by field with moon_paksha_2026_11_04: sealed holdout 20/20 and reported
offset 900/900 identical, metrics and summary unchanged.
Also records BUG-1575 (jyotish_engine MOOLATRIKONA gives the Moon Taurus
0-3; investigating, not changed).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
Data audit wave 2 (H) against PyJHora 4.8.6 and Swiss Ephemeris, 19 charts,
Raman and Lahiri passed explicitly to both sides.
- BUG-1566 (jaimini.py): the K.N. Rao Chara Scorpio / Aquarius lord read both
co-lords' dignity in the dasha sign itself, so Scorpio was always Ketu and
Aquarius always Saturn. Now one lord in the sign gives way to the other
(K.N. Rao; Rath p.43 (c); PyJHora); otherwise compared where each stands
(companions, Jupiter / Mercury / dispositor, dignity, modality, degrees).
MD years vs PyJHora 212/228 -> 225/228 (Raman), 226/228 (Lahiri).
- BUG-1567 (jaimini.py): K.N. Rao antardashas now run each MD's own
direction with the MD sign last (pp.35-36); book order 19/228 -> 228/228.
- BUG-1565 (gulika.py): Wednesday-Saturday night share table started one lord
early; Maandi and the four other time-segment points 79/114 -> 114/114
against an independent Swiss computation. Gulika unchanged (19/19).
- BUG-1568 (jyotish_engine.py): natal Panchanga vara / hora used the civil
date; a birth before sunrise belongs to the previous weekday (15/19 -> 19/19
vs PyJHora drik.vaara).
Narayana (legacy and Rath), transits, sunrise and day/night: no defect; the
second-cycle rule matches PyJHora's. Variants are listed in BUG_HISTORY.
jaimini.py is in the frozen identity: per ERR-110, new records frozen and
replayed as audit_h_2026_11_04; contract, freshness test and frontend gate
test rotated. Versus moon_paksha_2026_11_04: sealed 20/20 and reported offset
900/900 identical, metrics and summary unchanged. Chara sections of two
frontend fixtures and the rashi-dasha PyJHora fixture's native rows updated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
Data audit wave 2 (I). 04d901e1 (2026-06-27) renamed the evaluator's
pyjhora_* helpers to external_benchmark_* but left the rule expressions
calling the old names; the NameError is swallowed, so nine rules (Matru
Sneha on the parents card, Bhratruvriddhi, Amala, Sumukha, Bandhu Pujya,
Kaalanirdesat Puthranaasa, ...) never formed. The rule library now calls
the current names; logic is unchanged. Against PyJHora (Raman, mean nodes,
19 samples) the revived rules agree 18-19/19; the 20 Raman packet rules
go from 302/380 to 327/380.
_eval_parivartana resolved "$a" / "$b" with ctx.resolve, which knows no
any_pair bindings, so raja_yoga_parivartana (kendra lord <-> trikona lord
exchange, BPHS 34.11) never fired. Bare "$x" refs now go through
_resolve_binding.
New test scans every custom expression for names the evaluator does not
define (only parvata_yoga, non-classical and left off, is excepted).
Neither file is in the rectification frozen identity; sealed 20/20 and
reported-offset 900/900 replay identical to moon_paksha_2026_11_04.
BUG-1570 (Aries / Scorpio Venus malefic by BPHS 34.20 / 34.35) and
BUG-1573 (packet members with no evaluable detector, no D9 context in the
consultation yoga path) are recorded, not changed: the BUG-1570 trial cut
reported-offset top-1 by 3-5 pp and needs a product decision.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
The precise path doubled the Moon's Paksha in Kala (Raman, Graha and Bhava
Balas §55(c)) and then took that doubled value again as her Chesta (BPHS
27.18), so her Paksha entered the total four times; Chesta could reach 120,
clamping Kashta to 0 and pushing Ishta past 60. Every surveyed source
counts 2P.
Product decision 2026-11-04: BPHS (Santhanam) 27.10-11 + 27.18, as in
saravali.github.io, Maitreya8 (default) and jyotishyamitra. The Moon's
Paksha is the folded base, not doubled; waxing Moon benefic, waning Moon
malefic (BPHS 3.11, already what classify_drik_planets does, unchanged);
Chesta = Paksha. method_variants.moon_paksha names the convention. Other
planets, Drik and the legacy calc_kala_bala path are unchanged.
shadbala.py is in the rectification frozen identity, so per ERR-110 new
records were frozen and replayed as moon_paksha_2026_11_04. Compared field
by field with data_audit_2026_11_04: sealed holdout 20/20 and reported
offset 900/900 identical, metrics and summary unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
The first mahadasha starts before birth, so the nine periods end before the
120th birthday. Historical charts past that point had no covering period:
cmd_dasha's current_dasha was null and the consultation card's
timing.vimshottari was empty (Frida Kahlo at 2026-09-27).
next_cycle_mahadashas (chart_dasha/birth_moment.py) continues with the same
lords, order and full 365.25-day years until a period covers the reference
day; cmd_dasha, the chart surfaces and the consultation sub-periods share it.
Nothing is added inside the first cycle. The frozen rectification identity
(dasha_analyzer, domain_calculation_service) is untouched, so no re-freeze.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
Each domain house row on the consultation evidence card now carries
sav_strength: the raw SAV of the house's sign as 强 / 中 / 弱, never the
bindus. The band is the chart page's savBand (>= 28 强, 25-27 中, <= 24 弱),
sourced from references/ashtakavarga-complete-system.md §三 and the same
cut-offs as scripts/ashtakavarga.py sav_assessment (SAV_BAND_SOURCE).
A missing or mismatched engine row leaves the label off and lists a gap.
The system prompt's card description names the field in one clause.
Shadbala stays off the card (engines disagree).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
BUG-1545 (scripts/narayana_dasha.py, scripts/active_rectification_event_engine.py)
and BUG-1546 (scripts/shadbala.py) changed files in the frozen scoring
identity, so the integrity gate went red. Per ERR-110, new records were
created with each script's --freeze and replayed under the label
data_audit_2026_11_04; the scripts, contract, freshness test and frontend
gate test now point at them, following 129313e1. Prior records are kept
byte-for-byte.
Compared field by field with the d60_mapping_2026_10_29 records:
- sealed holdout: 20/20 trials identical; metrics unchanged.
- reported offset: 898/900 identical; 2 trials changed true_rank only
(one case, radius 60, offsets -30/-20: 76->67, 80->77). Top-1,
minute error, delivery coverage and summary are unchanged. Reverting
only the BUG-1545 files restores the old ranks; reverting shadbala.py
does not; a different PYTHONHASHSEED gives the same result.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
The chart API called calc_shadbala without a context and so used the
simplified fallback, and it rewrote `degree` as the in-sign degree, which the
precise Dig bala reads as a longitude. Readings and reports took the precise
path, so the same chart showed different strengths and a different strongest
planet (Jobs: Moon 10.61 vs 6.61 Rupa). The endpoint now builds the same
context as cmd_shadbala and passes absolute longitudes; a test locks the two
equal per planet. Also records the astrologer's ruling on waning-Moon Paksha.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
The natal marriage point was hard-coded to Venus - Saturn + Asc with no
zodiacal-order exception, while the annual Tajika point already followed
references/saham_rules.json (PVR Table 74: night reverses the operands, plus
the +30 exception). The same night-born person had two marriage points; 6 of
the 9 public charts move. The full reading, the vivah-saham command and the
consult card now share tajika._formula_saham_trace, with day/night from the
card's own day_night evidence. 0 deg Aries is no longer "missing".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
abs(Sun - Moon) / 3 was never folded to 0-180 degrees, so on charts whose
raw difference passes 180 benefics got more than 60, malefics went negative
and the doubled Moon passed 120 (Monroe's Moon -45.76; 5 of 19 benchmark
charts). PyJHora has the same line, which is why the comparison agreed.
Same audit, no code change: the +/-60 Kala gap is PyJHora's hora formula;
PyJHora is the outlier on Mars/Jupiter/Venus Chesta; Sun/Moon Chesta is a
school difference. Shadbala totals stay off the card until those settle.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
Ledger entry for upstream 5fb8379c with target 9d69cd56 and rollback list.
The progress note has the inventory, what was not ported and why, test
results, and the status of the two 2026-10-03 upstream issues. Both are
fixed upstream: KARAKA_8 now follows BPHS order, and Narayana profiles now
reach calc_narayana_mahadasha.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
The legacy current-period lookup took the age modulo the first-cycle total,
so anyone older than their first cycle (61-96 years on the nine public
charts) got a first-cycle period: an 80-year-old was placed in 2-12. The
second cycle runs the same signs for 12 minus their first-cycle years; the
lookup now places the age on both cycles and subdivides with the second
cycle's years. The rectification cache only covers the first cycle and falls
back to the live lookup past it, so first-cycle scoring is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
cmd_kp handed calc_kp_analysis the cusps from calculate_bhava_boundaries
(rounded to 4 decimals), degree_in_sign (4 decimals) and the chart's
display julian_day (6 decimals). KP star, sub and sub-sub boundaries are
cut on raw longitude, so a body or cusp sitting on a boundary could land
one cell off. The fix recomputes the JD from the civil instant with the
compute_chart_data formula, uses calculate_cusps directly, and uses
degree_raw % 30.
Only the KP part of 26d5a52d is ported. Its Vivaha Saham day/night change
and Yoga status pass-through are left for a separate decision. Test ported
from upstream tests/test_kp_unrounded_inputs.py.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
today_str and transit_reference_date already followed transit_date, then
today, then now. Narayana's current period read only args.today, and the
nakshatra-dasha age and the annual fallback age used datetime.now(). A
dated report therefore mixed two dates. The annual fallback age also used
days // 365, which moves to the next year of age early.
All three now take the same observation date. The fallback age counts
calendar birthdays. Tests are ported from upstream
test_full_reading_observation_date.py and test_tajika_report_age.py and
adapted to cmd_full_reading. Same-day web consultations are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
The full reading's inter-chart linkage appended D12 with a hard-coded None
ascendant, so every D12 row had house None even though d12_asc_idx was
already computed a few lines above. An unknown sign name also fell back to
Aries. Port upstream cf6c316d and f60568a0: pass d12_asc_idx through, and
raise on an unknown sign or a non 0-11 ascendant index. The full reading
records the error and the consult native layer marks it blocked.
Test ported from upstream tests/test_inter_chart_d12_house.py.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
The writing rules were only a slice of the agent's input; the rest pulled
answers toward a report. Product approved six cuts (2026-11-04):
1. Chat tool results no longer carry the router's English "Shared
mandatory baseline" (degrees, nakshatra, MEVG, case calibration, A/B/C
labels for every reading).
2. Checklist wording the model read aloud: house-counting is for judging
only; "双体星座" becomes the four sign names.
3. The bound skill excerpt keeps only the truth-boundary paragraph (KP /
Muhurta table, Ashtottari, P0/P1 list and three-layer method dropped;
marker kept).
4. The reference_transparency block goes: the card never carries it.
5. English rules that restated VISIBLE VOICE go; "keep conclusions
conditional" becomes "never call it verified, but say plainly what the
chart shows".
6. The rectification offer rule is the offer tool's description instead of
riding in every user turn.
Parents backtest: unglossed terms 16 -> 11 (std) and 25 -> 11 (verbatim
question); no absent parent written as present or involved.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
Product 2026-11-04: grounds may be listed in full as long as each one says
what it means; the "can't tell distance from friction" range is said once
per answer (merged when both parents are heavy). Backtest: answers repeating
the range 11/18 -> 6/18 (std) and 6/9 -> 1/9 (verbatim question); unglossed
terms 33 -> 16; still no absent parent written as present or warm.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
Backtest (parents, 9 Rodden AA figures, blind-scored) showed the slim voice
wrote absent parents as present and warm (severe 0 -> 8, then 1 -> 5) once
the presence range was gone. Product chose the hybrid: when a person's line
carries heavy pressure, give the range once in the model's own words and
write no portrait of that person's conduct or warmth on that side.
Hybrid round: severe 3 -> 0 (std), 2 -> 0 (correction), 4 -> 2 (verbatim
question); closing asks 26/26 -> 3/27; control false positives unchanged.
Also: correction turns say which line was off and never blame the user;
jargon and varga codes are said in plain words; house-counting is a named
ban; no mind-reading when the user asks for objectivity.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
The parents answer read as house-by-house fact chains with a gloss on every
house, opened with "both lines show pressure" instead of the pattern asked
for, and copied the example's 「实际是哪一种，告诉我一句」. The model was
satisfying ~13.9k characters of writing rules stated three times (voice,
English contract, user turn) plus copyable Good/Bad examples.
- VISIBLE VOICE is now the single statement: follow the user's own
structure, open with the overall picture, explain why each placement
means what it means, gloss terms briefly; then the red lines only.
- Contract and first/follow-up user turns point at it instead of
restating it; all example answers removed.
- AFFLICTION_RANGE_RULE keeps the red line (no biographical verdicts, no
reverse reassurance) but drops the forced range and closing ask; the
question's own premise counts.
- Contract tests cap the voice at 2600 characters and lock every red line
the old tests guarded; backtest script gains --question.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
- proxy.ts: no session cookie on `/` → 307 /login before any page code, so the
shell and 正在载入账户 no longer flash first
- login: segmented tabs fill the width and fit their track; six code boxes
span the form like the email field
- /rectification: drop the 先核对一下资料 step; every button sits in one row
at the card's foot instead of one inside and one below
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Empty or short input is one pill row (64px desktop / 56px phone, radius
999): 常用问题 | field | 40px send (44px target). Once the text wraps in
the single row or contains a newline the form gets .is-expanded: the
leading slot on top, the field at full width, send bottom right, radius 24
(22 on a phone). The field grows to 6 x 26px = 156px, then scrolls inside
(.is-capped, overflow-y auto) with a currentColor mask fade at the top
when scrolled (.is-scrolled). Focus tints the border and adds a 3px
action-soft ring. Send is canvas-muted / ink-tertiary while there is
nothing to send, action-fill once there is; stop is unchanged.
The switch is decided from the real textarea (composer-layout.ts):
expand as soon as the single-row content exceeds one line; collapse only
when the text has no newline and fits the single-row width with 16px to
spare (hysteresis). No change during an IME composition; it settles on
compositionend. Only classes on the <form> change and the DOM order
(leading slot, field, send) is fixed, so focus, caret and the preset
fill's （…） selection survive. Same component on the starter, in a
conversation and on the rectification surface.
Tests: new chat-composer-single-row-20261103.test.tsx (11); three existing
assertions updated with 原值/新值/原因. DESIGN.md composer section,
CHANGELOG and the pre-work ledger updated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
- Two half-width entry bars (76px, 72px stacked on phones): icon square,
title, one ellipsized line, right-side button. 生成报告 is the pale
accent (action-soft fill, action text, 600), 40px visual / 44px hit;
打开年运 is a bordered link. Phones show 生成 / 打开.
- 过往的报告: 共 N 份 only when every page is loaded; 刷新 removed, the
failed-read alert keeps 重试.
- 64px rows with hairlines; the whole row links to the report (generating
opens its progress); type icon (FileText / Sun / Heart), title, meta
(date · depth · themes, annual date only, failed rows show the failure
summary), plain status text by colour. ⋯ delete menu is a sibling of the
link. Phones drop the icon and the year.
- Signed-out / empty / failed states follow the same language; no large icon.
- Entry CSS moved from annual-report.css (off-allowlist 720px cut) to the
global 767 cut. Depth labels: DB only allows the four labelled values;
the raw "full" came from the screenshot mock, not a product bug.
- Tests: new reports-page-option-a-20261103; updated assertions carry
原值/新值/原因. DESIGN.md and CHANGELOG updated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
The not-started card no longer leaves its button row empty. That reserved
44px row now reads 「来源 <source>」 with the intake's short source names
(issue #144, birthTimeSourceOptions labels), and, when an earliest / latest
range was given, 「你给的范围 HH:MM–HH:MM」 beside it (wrapping to a second
20px line on a narrow phone, still inside 44px).
- rectificationDeclaredFacts: clock sources use the declared earliest /
latest, or the pre-#144 minutes either side of the stated time; period and
shichen use their clock range; 不清楚误差 gives the source only; unknown,
legacy_import and family_exact say nothing (the row stays empty).
- The range given is drawn as the band on the ruler (same pale-accent
style) with the reported tick on it, and the ruler's range rule treats it
as the window; the legend adds a band swatch with 你给的范围.
- Card heights are unchanged (244px / 234px measured, pending equal).
Tests: none state with a range, with 不清楚误差, and with no usable source
(unknown, legacy import); the lib cases for all six sources, the minutes
fallback across midnight, and the ruler including a 23:00–03:59 range.
The earlier none-state test assertion is rewritten with 原值/新值/原因.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
The /rectification status card follows the product owner's approved option A
(时间尺), with the requested change that the two buttons are separate:
- The card no longer repeats the person and 本人; the header chip shows them.
- Status row (还没校正过 / 上次没做完 / 已采用) with 已收 N 件大事 or the
done note on the right; one big tabular time; a ruler with the current
window as a pale-accent band, the reported time as an ink tick and the
adopted time as an accent dot; a legend line.
- The ruler range comes from the data (rectificationRuler): the window, else
the adopted or reported time, about half an hour each side, at least two
hours, snapped to half hours (whole hours past four hours); a window whose
end is before its start crosses midnight. Missing data hides the ruler,
band or tick; never NaN.
- In the card only the action on this record (继续上次 / 查看结果), bordered.
Below the card, on its own, one pale-accent button for a new case
(开始 / 新建一次 / 再校正一次). Full width on phones. No solid clay button
left on the page; the check step's 开始校正 uses the accent too and the
close-failed note sits beside it. The 新建一次 flow is unchanged (BUG-1516).
- Summary and pending cards share one measured height (244px desktop, 234px
phone) and the outer button row is reserved, so nothing moves when the
summary lands (BUG-1528). Message states use the same card and pattern.
- Tokens and color-mix only; 44px targets.
Tests: new rectification-entry-ruler-20261103 (ranges incl. 23:40-00:20 and
missing data, band/tick/dot positions, in-card vs outer buttons, no person in
the card, token and 44px rules); loading-shell card-size assertion rewritten
with 原值/新值/原因. DESIGN.md, CHANGELOG and the pre-work ledger updated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
Brings in the shell fixes (BUG-1532, BUG-1533). CHANGELOG, BUG_HISTORY and the
pre-work ledger conflicted only where both sides appended; both kept.
globals.css merged cleanly.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
- Popover collision: side "shift", align "shift", fallbackAxisSide "none",
padding 4 top / 8 elsewhere. Short of room (a window ~800px tall) the
panel shifts down inside the viewport and ends at the trigger instead of
falling back beside it. Measured in Chrome at 1280x800, 1280x860, 1440x900
and inside a conversation.
- Tests: the note only in its own domain, no hint line under titles, and the
caller's hide conditions (synastry, rectification, incomplete profile) and
first-blank selection unchanged.
- docs/testing/consult-preset-menu-20261102.md rewritten for the pill, the
two-column panel (hover, keys), the phone chips and pick-fills-not-sends.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
The 432px panel fitted neither above nor below the composer in a blank new
chat at 1280x860, so Base UI fell back to the trigger's right and covered
the input. The panel is now 404px (nav gap 0), and the Popover only flips
between top and bottom (fallbackAxisSide "none", collisionPadding 8). The kit
PopoverContent passes collisionAvoidance / collisionPadding to the positioner.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y
ClaudeandClaude Opus 5.5