Product updates for the app that turns your captures into organized research memory.
v0.19.0
v0.19.0
Public & live
Shopping Trips: discover, choose, and remember
Keep sale listings, your likely stops, a plan, and the day’s discoveries together. Your AI can save Facebook research and retrieve structured locations for mapping.
Created
Trips with photo cards, comparison table, personal selections and reasons, planning order, and a lasting research journal.
Original listing links, source dates, address precision, linked Lists and Missions, and private photos from you or your AI.
Five workflow tools let your AI save and retrieve trips, stops, and evidence with safe retries and corrections.
Fixed
Repeat source listings stay on one stop; AI recommendations remain separate from your choices.
Upgraded
Existing research sessions keep their IDs and captured items. Portable export includes Trips, research and linked shopping history.
What to check
Open Trips, add a sale, choose Likely, and review Your plan.
Open original Facebook links, save a journal note or photo, and export all stops for your AI.
v0.18.0
Public & live
Items first, shopping lists, and Places
Keep things you research, appraise, and buy in one Items workspace. Use Lists while shopping, and give each Place a useful overview, visits, and private photo gallery.
Created
Items brings Mission finds and appraisals into shared card/table views, with Appraisals and Purchased selections, filters, and sorting.
Shopping lists now have quantities and Need/Bought controls. Checking a line does not invent a price.
Places separates Overview, Visits, Items & appraisals, and Gallery. Upload private photographs of the place separately from photos of items found there.
Fixed
Home figures open their records; Items moves up in navigation and Queue sits beside AI Suggestions near the bottom.
Price comparisons keep pack sizes and currencies separate. Trip prices cannot disappear into an unrelated list, and a full history refuses new entries instead of deleting old ones.
Receipt confirmation selects one exact price, and concurrent place-description edits cannot overwrite an unseen revision.
Upgraded
Existing Item and Appraisal IDs, evidence, photos, and AI tool names remain intact.
What to check
Open an appraisal from Items, switch card/table views, and try Purchased.
Make a shopping list, set quantities, and check a line bought.
Open a Place, switch its tabs, and add a private place photo.
v0.17.0
Public & live
Lists you name, visits you keep
The rest of the antique-mall report. Lists you name yourself hold the things you buy again and what each cost every time, so your AI can say whether today's price is high or low and where it was cheapest. A visit is a real record now, and every place can carry your own long description.
Created
Lists: name your own — Recurring purchases, Tool items, Costco run — nest a dated trip inside a bigger list, and put things in them. Buying the same thing again adds to that thing's history instead of starting a new record.
Every list item reads its newest price against its own history — high, low, or about usual, with how many prices that stands on — and shows the cheapest places it was confirmed at.
Receipts: your AI can read one and record the lines as a trip. A price read off a receipt is shown dashed and counts for nothing until you confirm it; there is no way for an AI to confirm its own reading.
Visits: open one at a place, close it when you leave, and the appraisals and notes made on it hang together, so the next time you open the venue you see what you looked at last time.
A long description per place — layout, floors, consignment rules, discount patterns, which categories are strong — private to you even on a shared directory entry.
Fixed
The Visits column on the Marketplace Directory has said "nothing records a visit yet" since August. Something does now.
Upgraded
The connection catalogue is forty tools: seven new ones for lists, visits, and the venue description. Your AI can ask "what do I usually pay for this?" in one call.
What to check
Make a list, add a thing, record three prices at two places, and confirm the reading and the cheapest place appear.
Ask your AI to read a receipt into a trip, then confirm the lines on the trip page and watch them turn solid.
Open a venue, start a visit, add a note, end the visit, and confirm the note sits on the visit.
v0.16.0
Public & live
Built for the antique mall
A morning at a multi-dealer mall showed six small gaps between what happens in the aisle and what the record could hold. Each has a home now: which booth, that booth's discount, a reason to come back, whether you actually wanted it, and the tag price read after your AI already answered.
Created
An Appraisal can say why you asked — just curious, or actually want it — and the Appraisals index can filter on it, so the real leads never drown in the fun ones.
Booth or dealer number is a real field on Appraisals and venue notes, so "everything from booth 277" is a question the record can answer.
A venue note can carry that booth's standing discount and a come-back mark. The venue page shows the newest discount per booth and a Come back for list, and your AI gets both when it lists the place.
Add details on an Appraisal — the tag price you read after the answer, the intent, the booth — without touching the question, the answer, or what happened.
Fixed
Your AI can now check whether a place already exists by name in one small call instead of pulling fifty places and scanning them.
An asking price learned after an Appraisal was answered used to end up in a venue note. It now sits beside the answer where the outcome can be measured against it.
Upgraded
The connection catalogue is thirty-three tools: awb_update_appraisal is the details pass, and the venue and appraisal tools take the new fields up front.
Only the fields you send change. An omitted detail is left alone, never cleared — the same rule the outcome record follows.
What to check
Open an Appraisal, add a booth and mark it as wanted, then confirm the index's Why asked chips find it.
On a venue, note a booth discount and a come-back mark, then confirm both appear on the venue page and when your AI lists the place.
Ask your AI to add an asking price to an Appraisal it already answered, and confirm the answer and the outcome were not touched.
v0.15.0
Public & live
Your AI can work the new screens too
Appraisals, the Marketplace Directory, Suggestions and Items arrived as screens only you could use. Nine new tools let your AI work them alongside you — and the Mission desk now tells the truth about whether your AI reaches that Mission at all.
Created
Your AI can raise a Suggestion with its evidence, keep an Item's status current, remember what a venue was right and wrong about, and answer an Appraisal you snapped at an estate sale.
The connection catalogue is twenty-nine tools. The nine new ones read and write only your own private workspace; none of them touches the public places directory.
Fixed
A Mission desk used to show "Your AI · connected" with invented permissions whether or not anything was connected. It now answers the real question: nothing connected says so, a read-only permission shows exactly that, and a permission narrowed to other Missions says it cannot reach this one.
Nothing could reach the Suggestions screen before this, so it could never receive a suggestion from anyone.
Upgraded
There is deliberately no tool that accepts a Suggestion, and none that grows an Appraisal into a Mission. Those two decisions stay yours.
When your AI is refused a Suggestion it already asked about, it is told why — the rule, the reason you gave, and the one condition that would make asking again reasonable.
What to check
Open a Mission and confirm the AI card matches what you have actually allowed, including when nothing is connected.
Ask your AI to propose something, then open Suggestions and confirm the accept decision is still only yours.
Ask your AI to accept a Suggestion or grow an Appraisal into a Mission, and confirm it tells you it cannot.
v0.14.0
Public & live
Appraisals, venues, Suggestions — and a home to open on
The August design iteration. Signing in now lands on a five-slot home under one promise — assisting your AI to assist you with buying and purchases — and four new places hold buying work: a photo-first Appraisal, the Marketplace Directory, your AI's Suggestions, and Items with a life story.
Created
Appraisals: photograph one object, ask one question, and keep your chosen AI's answer attributed to it by name. The camera opens directly on a phone, and an estate sale with no signal still keeps the photo and the question until the network returns.
The Marketplace Directory is one list of places to buy, physical and online together. Venues you add are private, and your notes on a venue stay private even when the place itself is public.
Suggestions is where your AI proposes and only you decide. Accepting writes an editable draft that starts nothing and tells no one; declining records your reason and the one condition under which your AI may raise it again.
Items get their own index — the first read across every Mission — with what happened to each thing, where it is filed, and the numbered evidence behind each claim. Finding is still the neutral word for what research turns up; Item is what you see when you are looking at the thing itself.
Fixed
The public Sign in link led to a page that bounced straight back to the marketing page, so clicking it did nothing. Signed-out, that address is now the sign-in door.
Accepting a suggestion used to add to the one counted badge in the whole app — the no-pressure rule broken at the moment you followed it. Accepted suggestions now wait as quiet drafts, out of every count, until you release them.
An AI still waiting for your approval was offered an empty tool list, which every client reports as a broken server. It is now offered exactly one tool: a way to ask you for what it needs and hand you the link that approves it.
Upgraded
A labelled navigation rail with no icon-only mode, its own phone chrome with a raised capture button, and one browse bar — search, sort, filters, and a count line that says what it is hiding — across every index.
Your AI can attach one photo, under a permission you allow by hand, and Settings → AI shows "What your AI looked at" for every connection.
Three numbers the home design asked for have no source in this workspace, so the page says so rather than inventing them.
What to check
Sign in and confirm you land on the home page, and that every number there says what it leaves out.
Take an Appraisal photo on a phone with the network off and confirm it is still waiting when you come back online.
Open a venue and add a note, then confirm the note is yours alone even on a public directory entry.
v0.13.0
Public & live
Signing in is the approval, and you can take it back
Connecting your AI now works the way you would expect: you click Allow once and it has its everyday toolset for thirty days. One page lists every connection, what it may reach, and a Revoke that stops it whenever you decide.
Created
Settings → AI lists every connection with what it may do, when it last worked, how long it has left, and a Revoke that takes effect on that connection's very next request — no waiting for a token to expire.
Permission is a Buying record in plain language rather than something decided inside your AI's client. What no permission can ever allow — contacting a seller, spending money, deleting or exporting your records, or reaching anyone else's workspace — is stated as fact.
Fixed
People finished the connection, clicked Allow, and had granted nothing: Buying was waiting for a second approval on a page they had never been sent to, while their AI received an empty tool list that every client reports as "couldn't connect". Signing in is the approval now.
An unreachable sign-in provider used to be answered as "your token is bad", which sends a client off to fetch a new token from the service that is down. It is now an honest temporary failure with a retry.
A permission you revoked could be quietly replaced the next time that AI signed in. Two further ways to lose a revocation were found and closed alongside it.
The browser tab still carried the framework's starter icon. The whole brand — app icons, install icon, social card, and tab — is now Buying's terracotta.
Upgraded
Contributing to the public places directory stays yours to allow by hand. Everything that writes only into your own workspace comes with signing in, because your AI can correct its own work there.
The addresses your AI's client uses are never redirected, so a connection works from either spelling of the site.
All of this reached the live backend on 28 August. Until that deploy, the public pages described a boundary the running backend did not yet enforce.
Support moved to the shared desk at Assist With Life; Buying publishes no email address of its own.
What to check
Connect your AI, then open Settings → AI and confirm the connection is listed with what it may do and when it expires.
Revoke it, ask that AI to do the same thing again, and confirm it is refused immediately rather than after a delay.
Confirm that contributing to the public places directory still asks for your explicit allow.
v0.12.1
Public & live
Mission edits that respect newer work
Question drafts now conflict instead of overwriting a newer person or AI edit, criterion retries stay quiet, and the Mission desk states exactly how much recent history it loads.
Fixed
Opening the question editor refreshes its draft from the latest Mission, while the version is frozen for that editing session so a newer concurrent change produces a reviewable conflict.
Retrying an unchanged accepted criterion no longer adds a false update to Timeline or increments the Mission version.
The criteria rail no longer promises a proposal-rejection control that has not shipped.
Upgraded
Decision and source history now disclose their bounded recent windows instead of implying the desk loads every stored record.
What to check
Open a Mission question, change that Mission from another active session, and confirm saving the older draft asks you to review the newer version instead of overwriting it.
Confirm Timeline labels its recent decision and source-check limits and criteria guidance offers only controls that are actually present.
v0.12.0
Public & live
A Mission you can shape yourself
A signed-in person can now carry one private Mission from its question through criteria, notes or sourced Findings, and a bounded decision, then optionally hand the exact next step to their chosen AI.
Created
Each Mission desk now has a five-step person path for the question, accepted criteria, notes and Findings, decision, and optional chosen-AI continuation.
A person can save a note on its own or attach a public source with what that source supports or limits; both appear in the normal Finding and Timeline experience.
Person decisions are dated, attributed results with explicit selected, deferred, no-purchase, or needs-more-research outcomes and visible uncertainty.
Fixed
The Mission desk no longer stops at a read-only board after creation; the owner can revise the useful question and make the private workspace useful before connecting an AI.
Upgraded
Person-authored criteria and Findings receive stable keys and provenance so a later chosen AI can read, preserve, and safely continue the same Mission instead of rebuilding it in chat.
Decision revisions append to history rather than overwriting the earlier result, and the Mission Queue remains a bounded next-step handoff rather than expanded authority.
The human `/ai`, plain-text setup, and short/full agent guides now teach the same person-first Mission continuity contract.
What to check
Start a Mission, open its desk, and use Your Mission path to revise the question, add a criterion, save a note or sourced Finding, and record a decision.
Reload the Mission and confirm the question, criteria, Finding/source, latest decision, and Timeline provenance all remain visible.
Open Mission Queue and confirm it offers one bounded handoff to your chosen AI without implying seller contact, purchase, or broader authority.
v0.11.0
Public & live
The decision is visible when you return
A Mission Board now leads with the latest recorded decision, so a person returning after AI research can see the recommendation, rationale, uncertainty, attribution, and evidence path before studying the comparison.
Created
Each Mission Board now opens with the latest dated decision, including its outcome, rationale, remaining uncertainty, author, and selected Finding when one exists.
Fixed
A completed chosen-AI handoff no longer leaves its actual recommendation discoverable only after switching from the default Board to Timeline.
Upgraded
The decision snapshot links to the selected Finding and recent bounded source-and-decision history, while append-only earlier decisions remain stored instead of being overwritten.
What to check
Open a Mission whose AI research includes a decision and confirm the Board starts with that latest decision before the comparison grid.
Confirm the snapshot names remaining uncertainty and who recorded the decision rather than presenting it as Buying's own judgment.
Follow the selected-Finding and history links and confirm recent source checks and decisions remain inspectable.
v0.10.0
Public & live
One Mission handoff from question to saved result
A person can now hand one bounded Mission step to their chosen AI and see the sourced comparison, decision, provenance, and Queue outcome return to that same Mission desk.
Created
Each Mission now has its own Queue composer and bounded work list, using the same four person-facing states as the main Queue.
Mission-scoped handoffs tell the connected AI which Mission to read before research and preserve person and AI actions on the Mission timeline.
Fixed
Normal Queue entry now preserves Mission and optional Finding scope through the offline-safe capture path instead of losing that context at sync time.
A claimed Mission handoff cannot be completed into a different Mission, preventing a valid lease from attaching research to the wrong decision.
Upgraded
The complete-result save now joins the Queue result, Mission comparison, sources, decision, timeline provenance, and normal web view in one bounded workflow.
A fail-closed exact-subject cleanup can remove the marked joined Mission, Queue activity, MCP operations, and structured result after soft-launch proof without touching unrelated research.
What to check
Start a small Mission, open its Queue tab, and add one specific research or comparison handoff.
Have your connected AI pick up that handoff and save its result, then confirm the Mission Board, Timeline, and Queue agree.
Open the Queue item's activity to see the person handoff and the chosen AI's attributable completion without exposing its private claim token.
v0.9.0
Public & live
A clearer first Mission
Assist With Buying now gives a signed-in person a direct, durable Mission starting point, keeps earlier research distinct, and uses truthful product identity and connection guidance throughout the ordinary journey.
Created
Start a typed Mission directly from the Missions page with a name, useful outcome, and the view that best fits the buying goal.
Fixed
The signed-in shell, browser titles, install metadata, sharing attribution, and public pages now consistently use the intended Assist With Buying name.
Clearly marked workflow-evaluation records no longer appear as ordinary saved research; stored data is retained for a separately reviewed cleanup.
Public guidance no longer promises self-service deletion, granular AI grants, universal client compatibility, or revocation proof that the current product does not provide.
Upgraded
Canonical Missions are now the primary desk, while pre-Mission records remain available under the honest Earlier saved research label until migration is authorized.
AI setup now distinguishes the current Codex production receipt from unproved setup guidance for other clients, and the tool count comes from the released catalog.
Every release note now carries an explicit state so a reader can distinguish work in review from what is public and live.
What to check
Sign in, open Missions, and start a small Mission without going through Capture or Queue.
Confirm the shell and browser title say Assist With Buying and earlier research is labeled separately.
Open /ai, /privacy, and /updates and confirm the connection, deletion, and release-state boundaries agree.
v0.8.0
Public & live
Your AI can now build a durable Buying decision
Assist With Buying now gives a connected AI one complete Mission workflow: it can retrieve the person's Buying context, save products and offers separately, compare options against explicit criteria, preserve dated sources and prices, record a recommendation, and return the result to the normal Mission experience.
Created
A Mission-native 18-tool connection lets a chosen AI brief, search, read, and save the durable records that make up a real purchase decision.
One complete-result save can preserve products, Mission-specific options, dated observations, comparison criteria and cells, source outcomes, and a decision together; granular tools remain available for corrections.
Mission timelines now show the AI's source checks, recommendation history, uncertainty, and attributed activity instead of leaving the useful research inside one chat.
Fixed
Clerk OAuth access tokens now authenticate correctly while browser-session tokens and wrong-resource tokens still fail closed; Codex production acceptance verified this path, while other clients remain separately unproved.
A newly signed-in person with no workspace data now sees Queue's useful empty state instead of a blank body.
Upgraded
Mission reads and searches are bounded, writes are replay-safe, evidence history is append-only, and workspace identity is derived on the server rather than accepted from a tool call.
Agent setup guidance now documents the exact current catalog, the one-call workflow, the privacy boundary, and what remains Partial or Later.
A removable, exact-account acceptance path proves production MCP saves can appear in the signed-in Mission UI and be completely cleaned without touching another workspace.
What to check
Connect a supported AI from /ai and confirm it discovers 18 Assist With Buying tools after browser OAuth.
Ask the AI to research a purchase Mission and confirm its options, criteria, source checks, prices, and recommendation appear together on the Mission board and timeline.
Confirm the AI describes unsupported seller contact, offers, payments, purchases, and destructive changes as unavailable rather than implied permissions.
v0.7.0
Public & live
The Queue becomes a real handoff desk
Queue is now a complete, Buying-native place to leave work for your chosen AI, see who or what it is waiting on, answer a blocker, inspect attributed activity, and return to a readable result without treating Queue like an autonomous task runner.
Created
A four-state handoff ledger makes Needs you, Working, Waiting for your AI, and Done visible at a glance on phone and desktop.
Each Queue card now keeps the directive, requested outcome, exact next step, authority boundary, actor activity, and result together as one durable record.
A compact directive composer lets you leave the useful outcome directly in Queue, with full photo and product-detail capture still one link away.
Fixed
Waiting for your AI now says plainly when Buying has not observed an AI tool call and never suggests that saved work has started on its own.
Canceled handoffs remain readable inside Done and can be restored, while Needs you answers return the handoff to Waiting for your AI.
Queue now has purposeful loading, filtered-empty, whole-page error, and save-conflict recovery states instead of silent gaps.
Upgraded
Activity opens only when requested, loads in bounded pages, and shows the actor, time, explanation, and Queue version without exposing coordination secrets.
Working cards show the current holder, lease boundary, and bounded attempt count; Done cards distinguish a result, cancellation, and expiry without adding a fifth state.
The authenticated Queue journey is now covered by a removable non-production proof path, while production data remains outside fixture testing.
What to check
Open Queue on a phone and desktop and confirm all four states remain visible without horizontal scrolling.
Open one handoff's activity and confirm the history clearly says who acted, what changed, and when.
Check that a waiting item says nothing is running automatically, a Needs you item accepts an answer, and a canceled Done item offers Restore to Queue.
v0.6.0
Public & live
A Queue your AI can safely pick up and hand back
The Queue now gives people and their chosen AI one durable handoff record: clear human-facing states, guarded pickup, bounded recovery, questions that return to the person, and results that remain linked to the work that produced them.
Created
A complete four-state Queue — Needs you, Working, Waiting for your AI, and Done — shared consistently across Queue, Capture, Sessions, and AI guidance.
Durable Queue handoffs that carry the requested outcome, current step, priority, authority boundary, linked context, result, and attributable activity across AI sessions.
Guarded multi-AI pickup with one-winner claims, renewable work leases, version checks, and private claim tokens so two assistants cannot silently complete the same handoff.
Fixed
Expired work leases return items to Waiting for your AI instead of leaving them shown as Working indefinitely.
Automatic retries stop at their stated limit, while a person's answer, revision, or restore deliberately opens a fresh bounded attempt window.
Lease recovery now keeps linked shopping-session status in agreement with the canonical Queue item.
Upgraded
The AI tool contract now supports paginated Queue discovery, explicit claim and heartbeat transitions, blocker questions, bounded failures, release, completion, and activity history.
Completed work keeps structured result links and provenance, while cancellation stays internal instead of becoming a fifth person-facing Queue state.
Agent-readable documentation and portable exports now describe the same Queue lifecycle and continuity fields as the application.
What to check
Open Queue and confirm the four state names match Capture and Sessions.
Check that a Needs you item explains the decision required, and that a completed item links back to its saved research result.
Connect your chosen AI and confirm it can claim one waiting handoff without implying that Buying runs an AI itself.
v0.5.0
Public & live
The terracotta rebrand, and the road to Missions
Assist With Buying now looks like what it is — one space in the Assist With Life family, in Buying's own terracotta — with a rebuilt homepage, a reorganised app around Missions, Queue and the Library, and a foundation of privacy and honesty fixes. The full Mission model (typed Missions, the criteria board, Finding records) ships in this codebase but stays dark until the data migration turns it on — the homepage roadmap tracks it as Next.
Created
A rebuilt public homepage in the family's language, with the research-loop library your AI can read at /loops.txt.
Missions and Queue as the app's spine — Missions is the signed-in landing page, the Queue is a first-class screen, and Capture is one button that follows you everywhere.
Card and table views on Missions and the Library, with honest counts that say exactly what a filter hides, and saved views you can name, apply and delete.
A new brand mark: the terracotta shopping-bag square with the stacked wordmark, at every scale from the homepage to the mobile bar.
Fixed
Every count now states what it excludes — a needs-you strip that previews five items says 'showing 5 of 9'.
Derived price summaries refuse to blend currencies or invent a dollar sign for rows that never stated one.
Upgraded
The site moved to assistwithbuying.com on 28 July, the old address forwards every page to it, and the sign-in change that came with the move asked everyone to sign in once more.
Seller contact details never leave the server on any ordinary read or portable export — screens and exports carry only a presence marker.
Demo and example data can only be written on an explicitly non-production deployment; the server refuses everywhere else.
Generated API bindings are content-pinned and provenance-verified against real code generation.
What to check
The new homepage, and the Missions landing page with your real research in card or table view.
The Library's saved views, and the table views' honest sorting and count labels.
The homepage roadmap: everything marked Now is live in this release; everything Mission-shaped is marked Next.
v0.4.0
Public & live
Shareable research, under your control
You can now turn one private dossier into a deliberately selected, read-only research page for someone else — with a live preview, optional expiry, immediate revocation, and no accidental spillover from the rest of your account.
Created
A Share research flow on every dossier with individual photo and section choices, a safe profile-summary switch, and a separate value-range opt-in that starts off.
Three recipient presentations — Field journal, Gallery, and Quiet archive — using the exact same live projection as the owner preview.
A Shared by me control center with lightweight view counts, expiry, live edits, safe link replacement, per-link revocation, and revoke-all.
Fixed
Unknown, expired, replaced, and revoked links now converge on one generic unavailable response, including selected media requests.
Shared pages and photos are server-rendered or transformed on demand with no-store and noindex controls, so revocation is not undermined by a long-lived page cache.
Upgraded
Capability tokens use 32 random bytes and are stored only as SHA-256 digests; AWB cannot recover an old raw link after creation.
Portable exports now include share-management records while still excluding media bytes and recoverable raw capability tokens.
Public privacy, security, terms, and AI-facing guidance now describe the dossier-sharing boundary precisely.
What to check
Open a dossier, select one section and photo, and confirm the live preview exactly matches the recipient page.
Edit that share from Shared by me and confirm the existing recipient page changes without a new link.
Revoke the link and confirm both its page and selected image become unavailable immediately.
v0.3.0
Public & live
The field-journal redesign
The whole app moves from its MVP look to a real design: a warm paper-and-ink theme with brass accents, serif display headings, a proper navigation shell on desktop and mobile, and every page rebuilt on one consistent layout system.
Created
A desktop sidebar with archive search, navigation, your AI connection status, and account controls.
A mobile shell: slim top bar plus bottom tabs with a raised capture button.
A real landing page for visitors, and a signed-in overview desk showing recent research, queue counts, active sessions, and AI status.
Fixed
The app body was rendering in Arial instead of the intended typeface; typography is now deliberate everywhere.
Dark mode now comes from one shared color system instead of hundreds of per-element overrides, so light and dark stay consistent.
Restored the missing backend functions on production (public locations and workspace search had been failing since the July 5 release).
Upgraded
Library, Sessions, Search, Places, and What's New pages rebuilt on a shared page-shell pattern.
Capture, AI setup, dossier, and session detail pages restyled into the new design without changing how they work.
shadcn/ui component library and lucide icons added as the foundation for future UI work.
What to check
Sign in and confirm the sidebar appears on desktop and the bottom tabs appear on your phone.
Open a dossier and a session and confirm everything still reads clearly in both light and dark mode.
Open /locations and /search and confirm they load without errors.
v0.2.2
Public & live
Value-write correctness fixes
A small correctness patch on top of the value-evidence ledger: AI-supplied value data now validates more carefully, a value range can carry its real "as of" date, and the release log now surfaces its own "What to check" notes.
Created
The What's New page now shows a "What to check" section for each release.
Fixed
Value ranges from AI agents can now record a real date again (the "as of" timestamp was being rejected).
AI value evidence is now rejected if its headline amount falls outside its own min/max range.
Removed a duplicate currency check that always reported the wrong field name, so validation errors are clearer.
Long auto-generated value notes are now length-capped to match the rest of the system.
Upgraded
Fixed the project type-check so it no longer trips on the value-write test file, keeping the safety gate trustworthy.
Added regression coverage proving a value range's "as of" date round-trips correctly.
What to check
Open /updates and confirm each release now shows a "What to check" section.
Confirm AI-saved research still records value ranges and used/new evidence normally.
v0.2.1
Public & live
Release safety checks
This patch strengthens the behind-the-scenes checks around AI-assisted research saves and value evidence. It keeps the visible value-ledger release stable while adding guardrails for the next cleanup and harness work.
Created
Added focused contract checks for how AI research saves item targets, dossier sections, and value evidence.
Added a static value-write guardrail that verifies manual and AI value evidence stay aligned around amount, range, source, market, similarity, actor, and timestamp fields.
Fixed
Resolved the release-branch stacking around the agent save helper work so the focused safety slices could land on main cleanly.
Kept the true Convex function harness follow-up separate instead of claiming runtime persistence proof that is not shipped yet.
Upgraded
Improved release confidence for the value-ledger workflow without changing the user-facing value entry UI.
Prepared the next TPP-48 / GitHub issue #7 harness work with clearer boundaries and proof expectations.
What to check
Open /updates and confirm v0.2.1 appears above the v0.2.0 value-ledger release.
Use the value-ledger UI from v0.2.0 as the product-facing release to review; this patch is mostly safety and release-readiness work.
v0.2.0
Public & live
Value evidence ledger
Research items now use a shared value ledger so used and new value evidence can be recorded, summarized, and reused by the app and AI agents without keeping a second old price system in the UI.
Created
Added USED and NEW value summaries on research items with a compact Add new Value flow.
Added value evidence fields for amount/range, condition, discovery method, market family, similarity, evidence link, date, and notes.
Connected capture, research follow-up, and session surfaces to the same value-ledger model.
Fixed
Removed the old user-facing compare-against-price editor from the clean release path.
Kept old price rows as compatibility fallback data instead of a second visible value system.
Upgraded
Updated the AI/MCP guidance so agents write structured priceValues rows and treat priceContext as legacy fallback only.
The value data contract now supports append-only evidence rows with who/when metadata and similarity context.
What to check
Open a research item and confirm the USED/NEW strip and Add new Value panel feel compact and useful.
Check capture and follow-up edits for the shared Value evidence fields.
Confirm older value data still displays as fallback evidence without exposing a separate old price editor.
v0.1.0
Public & live
Initial release-log baseline
Assist With Buying now has a standard What's New page for tracking product changes. This first entry is intentionally conservative: it marks the release log itself and the already-visible app surfaces, while leaving unfinished work out until it is ready.
Created
Added the standard What's New page at /updates for future product release notes.
Established a first baseline for the existing capture, research, sessions, and AI-connection surfaces.
Fixed
The /updates route now loads as a real page instead of returning a missing-page response.
Upgraded
Release notes now follow a consistent Created, Fixed, and Upgraded structure.
Future entries can separate shipped product changes from in-progress research and experiments.
What to check
Open /updates and confirm the baseline is clear, user-safe, and does not overclaim unfinished work.