Repository navigation
Comparison research can now be saved with its original evidence, checked answer
or failure and actual model inputs. Reopen without fetching data, rerun frozen
inputs or refresh current sources as separate linked runs. Explicitly assess
saved headlines and use selected retained opinions in a new Deep interpretation.
Text tone and asset implication remain independent; failed methods stay failed.
Standalone headline evidence defaults to 90 days from its latest recording;
saved runs protect their referenced evidence. Set NEWS_EVIDENCE_RETENTION_DAYS=0
to disable expiry. Home also reconstructs approximate 30-day portfolio history
from the ledger, with legacy adjustments available in a compact disclosure.
Implementation is merged through PR #283. Release-prep PR #284 passed all three
CI jobs and merged at 61599bc. Upgrade applies additive
Alembic migrations 0019_news_evidence and 0020_comparison_captures; Docker
runs Alembic before startup, while existing local installations must upgrade.
No dependency, provider or embedding change is required.
Added
- Retained headline lifecycle — standalone headline/assessment bundles default
to 90 days from their latest recording, with configurable expiry (0disables).
Saved runs protect referenced evidence until their last reference is removed.
History shows eligibility/protection and supports confirmed owner deletion of
unreferenced bundles. Bounded daily cleanup preserves saved briefings and checks
references under the writer lock; no migration or stored-payload rewrite. - Deep interpretation of retained sentiment — select up to five retained
headline observations from a saved news briefing and save a separate linked
interpretation using frozen market evidence. Inspect original text, independent
methods, scoring dates/inputs and checked citations; failures remain failures.
No source refresh or rescoring runs. - Reconstructed portfolio history — Home replays the primary portfolio's
ledger across the last 30 UTC days, including cash, closed positions, fees and
dividends. New purchases no longer reset the window. A daily inspection table
and Portfolio HQ value sparkline expose the approximate backfill; the main
comparison adjusts deposits/withdrawals as end-of-day flows. A new owner-only
JWT history endpoint provides the same values. Missing records or valuation
inputs remain explicit gaps; no database backfill or migration is required. - Headline sentiment in saved briefing review — explicitly select up to ten
saved headline references, retain assessments with cancellable progress, and
inspect original text, methods, scoring times and model inputs. Keyword tone
and Fast asset implication stay separate; failed models are not shown as
neutral. Paginated retained history works after source-briefing deletion. - Durable selected-headline assessments (API) — explicitly score up to ten
headlines from an owned saved comparison, retaining exact text and independent
keyword tone/Fast asset-implication assessments. Model failures stay failed;
cache hits preserve original scoring times and model inputs. Owner-only reads
work after source-capture deletion. No migration or legacy-score backfill. - Linked saved comparison follow-ups — Re-run saved inputs reuses the last
dispatched prompts, response format and generation settings with the current
Deep model and frozen evidence, making one checked completion. Refresh with
current data collects new evidence for the saved selection. Both save a
separate run with inspectable lineage; deleting a source keeps its children
usable. Progress and cancellation preserve the original and avoid automatic
retry. No migration or provider/configuration change is required. - Saved comparison briefings in Market Context — choose Generate and save
briefing to retain a fresh run with its evidence, answer or typed failure and
model inputs. Review and paginate your saved runs, inspect their original
citations and inputs, or explicitly delete them. Reopening performs no new
market or model work. A cancellable progress endpoint acknowledges saving only
after commit; interrupted acknowledgement can be checked in saved history.
Temporary generation remains available. - Explicit durable comparison captures (API) — authenticated
POST /api/market-context/briefing-capturesgenerates from fresh server-fetched
selections and retains the full bounded evidence, checked answer or typed
failure, public model profile and dispatched model inputs, including repair
and structured-format/token retries. Owner-scoped list/reopen/delete endpoints
recover the original record without fetching sources or running inference.
Additive migration0020_comparison_capturespreserves existing research.
Existing generation endpoints stay ephemeral.
Fixed
- Compact portfolio history notices — show one collapsed legacy-adjustment
summary, with transaction IDs, quantities and reconstruction explanations
available on demand. Active history failures and other warnings remain visible. - Legacy opening portfolio positions — reconcile legacy add/remove inventory
with today's holdings by estimating opening quantities before the 30-day
window's recorded trades. Home and the API expose the original quantities and
non-cash adjustments. Keep all in-window activity, reject negative opening
balances/remaining oversells and leave stored holdings/transactions unchanged. - Legacy add/remove portfolio history — exclude unmatched sales before the
displayed 30-day window, including their proceeds and fees, rather than blocking
reconstruction. Warnings identify every excluded row and explain possible cash
differences from current accounting. In-window oversells and remaining quantity
mismatches outside the legacy opening adjustment still fail explicitly;
no stored transaction is changed. - Benchmark prior close at a new holdings window — retain earlier benchmark
observations when the current basket starts after the last market session,
such as a weekend. The response includes quote as-of dates and coverage bounds;
Home shows the actual close used and diagnoses missing overlap separately from
missing history. Very short histories return null alpha/tracking error instead
of NaN or invented zero metrics. No portfolio/accounting migration is needed. - Portfolio benchmark daily alignment — normalize daily provider timestamps
to UTC dates before portfolio/benchmark alignment. Crypto midnight bars and US
equity session bars no longer lose their overlap and falsely report benchmark
history unavailable. Hourly timestamps and genuine missing-history failures
retain their behavior; no benchmark/provider substitution or accounting change. - Home portfolio history — remove the hard-coded 24–26 million demo curve
and clear old data on failed refreshes. Resolve the actual primary portfolio
ID, use only shared observations in the last 30 UTC calendar days, and rebase
holdings/benchmark to 100 at the same first date on one shared vertical scale.
Show the actual dates, base currency, history status and missing-input reasons.
The initial current-open-holdings curve is superseded by the ledger history
described above; Portfolio HQ now uses its value history. Overlapping polls
and late prior-scope responses cannot replace data. - BTC hourly candles and inspection — fold Yahoo's off-boundary, zero-volume
flat quote tail into an existing crypto period candle, preserving its open and
volume while updating close/extremes. Omit a quote-only tail when its period is
missing instead of inventing a period open. This normalization excludes
CoinGecko and non-fixed weekly/monthly intervals. Hovered OHLCV now takes
precedence over the latest candle, with volume from the same selected candle;
leaving the chart restores latest values and scope changes clear inspection. - Stock-chart candle refreshes — accept OHLCV and session corrections when
candle timestamps and row counts stay unchanged. Full history can now replace
a flat quote-only candle. Live aggregates starting mid-candle preserve the known
open/range and do not erase known volume with a zero sample or sum overlapping
volumes. Identical history retains stable references to avoid render loops.