Skip to content

Latest commit

 

History

History
225 lines (219 loc) · 456 KB

decisions.md

File metadata and controls

225 lines (219 loc) · 456 KB

dialed.run — Decision Log

Reconciliation decisions from the design/plan merge, 2026-08-30. Each entry: what was decided, what it overrode, and why. Referenced from other docs as D-NN. New decisions append here.

# Decision Overrode Why
D-01 MVP is social-first; the call is post-MVP (first epic). Design Milestone 1 ("prove the loop" = the call). Owner call. Feed + logging habit ship first; every schema stays recommender-ready so the call lands without migrations.
D-02 Superseded by D-159. Hand-validation study runs in parallel, non-blocking (task 090). Design's "validate before building the engine" hard gate. Infrastructure is needed regardless; findings tune the call epic.
D-03 TanStack Start replaces Hono JSX + htmx. Backend (CF Workers/D1/R2/Queues, $5 plan) unchanged. Plan's "htmx, no SPA framework — fixed". Design's app-like flows (sheets, multi-select, animation) exceed comfortable htmx; TanStack's end-to-end typing suits parallel agents; CF support is official (partner, framework guide, active investment). Youth risk accepted.
D-04 No comments at launch. Design D screen (comment thread) and E1 comment counts. Solo-operator moderation burden; the plan's argument stood. Revisit post-launch (post-MVP epic).
D-05 Verdict is per-run, 5-state, stored int −2..+2; optional run tags + per-item flags (too_much/not_enough) + private notes. Plan's per-item 3-state verdict; design's open 3-vs-5 question. One-tap logging (≤15s target) with finer scalar signal; item flags keep the item-level data the plan wanted without a survey.
D-06 Garment model = category × layer + typed attributes; zod discriminated union on category at the contract layer over a flat nullable D1 table. Plan's 9-value body_part enum; design's 6 stored categories. Splits "where it goes" from "how it stacks" (recommender needs layer); design's closet groups become derived UI groupings; union gives per-category type safety while filters stay indexed SQL.
D-07 V1 is online-required. Nothing may preclude a later PWA/offline layer. Design flow map's "offline is normal" (queued logging, cached closet). Feed-first MVP weakens the 5am-basement case; offline is a named post-MVP epic, not silent debt.
D-08 V1 capture = tap-list + forms. No vision model in v1. Design O2/F auto-read attributes. Vision capture pairs with the call epic; tap-list (curated per-climate lists) approximates the onboarding speed cheaply.
D-09 Kits deferred to the call epic (no kits tables in v1 schema; expand migration then). Design's pervasive kit surfaces. Kits earn their keep as the call's output and the feed's conversion moment — both deferred.
D-10 E2 ships lite: consensus block only; stranger cards, follow CTAs, full ranking deferred. Design's full E2 as default tab. Keeps the differentiator and cold-start value at a fraction of the scope and moderation exposure.
D-11 Reaction is "useful", single kind. Plan's likes. "Reference, not performance"; better future recommender signal; free at schema level.
D-12 Verdicts are public on shared posts (full 5-state label). Aggregate history, coverage, item flags/notes, thermal level: never public. Design's self-contradiction (badges on E1 vs "verdicts stay private"). A post without its verdict isn't useful reference; "raw comfort data" = the aggregates, which stay private.
D-13 Bulk GPX import + weather backfill + memorable-run tagging cut from v1; re-evaluated at the call epic as "conditions distribution + optional extreme tagging". Design O4 ("highest-leverage screen"). With the call deferred, seeded history feeds nothing; retro-tagged verdicts are noisy; cutting it dissolves the Visual Crossing backfill-budget problem.
D-14 Strava reminder webhook stays in MVP. No activity data stored, ever. Design's GPX-only intake framing. The reminder is the retention hook that makes people log at all.
D-15 Onboarding v1 = O1 (thermal level + location + units) + tap-list, own lane (105), after closet merges. Design's six-screen onboarding. O1 and the tap-list are cheap data collection that v1 features (consensus, coverage chart) and the future call both want.
D-16 Five tabs kept; Call tab is a coverage-ladder teaser until the call epic. — (new screen; design delta). Turns the gap into a logging-motivation loop instead of a dead tab.
D-17 Profile v1 = stats, temperature-coverage chart, most-worn, recent entries. No kits row, no calls-dialed %. Design G in full. Coverage is buildable from v1 verdicts and becomes the call's progress bar later.
D-18 Discovery deferred; v1 = username search + follow from posts/profiles. Design I (overlap ranking). Overlap math needs the cohort machinery; deferred with it.
D-19 Sharing default: public, per-entry toggle at log time, per-user global default (can default private). Private entries invisible to feeds and aggregates. Plan's all-public; design's explicit opt-in. Maximizes reference density while keeping the honest-logging escape hatch.
D-20 Guardrails enforcement loop kept (stop gate, fixer agents, commit gate). Prerequisites (release tag, live-loop proof) are workflow items; htmx-specific lint rules replaced by typed-route/no-string-URL rules. — Owner's existing investment; the parallel-agent model depends on machine enforcement.
D-21 Unified docs live at repo root as the real structure (this doc set). design/ and plan/ are read-only archives. plan/README's promote-later caveat. Agents start from the real thing; no promotion step.
D-22 Design-doc review is async (batched): agents commit the design doc and proceed; human comments are interrupts. Schema/bindings changes still hard-stop. PRs remain the hard gate. Plan's design-doc hard stop. Owner's attendance preference; packets + gates carry the risk.
D-23 T&S floor (now task 106) unchanged as launch gate, minus comment moderation. Private dogfood allowed pre-gate. — Confirmed as planned.
D-24 Weather provider stays Visual Crossing behind WeatherProvider. Manual-temp fallback allowed when unresolvable, stored source='manual', excluded from aggregates/training. Indoor runs skip conditions. Design's "NWS" label; brand principle 02 read literally. Adapter seam already exists; principle 02 means "never required", not "never possible".
D-25 Recommender-ready seams built now, silently: runs.effort (nullable), users.thermal_level, garment attributes + est_temp_*, metric storage, per-band verdict counts. No other speculative abstraction. — The cost is columns; the payoff is a migration-free call epic.
D-26 Canonical products table in v1. Garments link (nullable FK) to shared product rows created create-if-missing on normalized brand+name; curated seed list of running brands backs autocomplete/normalization. Text-only brand/name fields. Specific products behave differently (recommender) and showing off specific gear is core social appeal; text identity can never make "worn by N runners" honest. Product names are UGC — 106 extends report→review to them.
D-27 Capture is identity-first. Amended 2026-10-10 by round 35 #52b: F now asks Category and Type first, then Brand and Name, at every width, because the type picks the glyph and the defaults. A generic entry is still one submit away, and the rest of this row stands. Add-garment leads with brand (autocomplete) + model name; tap-list rows are explicit [GENERIC] placeholders, upgraded during onboarding (P2.5: "make the ones you reach for real"); closet shows a quiet generic-count nudge. Revised F's name-first form; tap-list rows as permanent generic records. The v1 cut of vision capture + link paste had made generic the path of least resistance; generic must be a scaffold, not a dead end.
D-28 Brand/product identity renders everywhere it exists — feed cards, post-detail kit rows, picker chips, most-worn (text only, no logos/links, per the honesty rule). Generic items render as category nouns; the contrast is the enrichment pressure. — The social show-off value the owner named; original D screen always assumed it.
D-29 Social-proof ownership counts derive only from public entries in v1 ("worn by N runners"), never from closet contents. "In your follows' closets" waits for an explicit closet-visibility setting. Original D copy ("in 3 of your follows' closets"). Settings copy promises closets are private; small-N follow counts make closet-derived counts identifying.
D-30 Product dedup = normalization + autocomplete steering; no merge UI in v1. 106's admin page gets a read-only duplicates report; merge tooling is a post-v1 micro-task. — Accepts some dirt; merge (FK re-pointing) isn't launch-critical.
D-31 Product intelligence pipeline ships in v1 (lane 107). Pasted URL → fetch ladder (JSON-LD/Shopify product JSON → OG tags → LLM extraction) → typed core columns + raw extracted JSON + full-page snapshot and primary image to R2. Retention-first: links rot, snapshots let extraction re-run and improve later. Fabric composition is a first-class extracted field. D-08's "no extractor in v1" (partially reversed — vision capture stays deferred). Deep product data (composition especially) is the long-term moat; most running brands are Shopify (structured data, no AI needed for the majority); the LLM rung is one queue consumer, same shape as the import pipeline.
D-32 Extraction model behind an adapter; 3-way eval decides (Claude Haiku 4.5 vs GPT-5.6 Luna vs DeepSeek V4 Flash on Workers AI — the third slot amended 2026-08-30 from Llama 3.1 8B: Workers AI now hosts current-gen DeepSeek/Qwen models, so the on-platform/free slot no longer has to be the weak one) on 20 fixture pages in 107's design phase; GPT-5.6 Luna presumptive default (amended 2026-08-30 — initially Haiku 4.5: comparative benchmarks put Luna ahead 13/15 with a higher intelligence index, it's a newer generation, explicitly positioned for extraction, and 4–5× cheaper; the single-vendor argument couldn't carry a deficit on both quality signals and price). Cost immaterial at launch scale ($1–25/mo across all options); decision reversible by re-running over snapshots. — Quality and operational fit decide, not price; measure rather than argue priors — and update the presumption when the evidence moves.
D-33 Strava notifications carry zero activity data. No distance/pace in the notification body; deep-link goes to +Add, not a pre-created run. Revised L screen's mock ("6.4 mi this morning", "run pre-created with conditions"). The revision violates the non-negotiable compliance rule; the design's own L1 copy ("we store nothing") is the correct spec.

Appended after reconciliation

Numbering: decisions are D-NN (D-01, D-02, …). The register, docs/deferred.md, is R-NN; it used D-N too until 2026-10-02, when it took its own prefix and kept its numbers.

# Decision Overrode Why
D-35 dialed.run has an admin section: The Desk, at /desk. One route behind the existing admin check, its own shell, always dark whatever the operator's theme, hi-viz as its only accent, desktop-first, and never linked from the runner app. Its destinations are Today, Review, Duplicates and Runners — banning lives on a runner's page rather than being a destination, and the digest is Today, with the email being Today sent to you. Lane 106's working assumption that each operator surface would be a known URL, defensible for one screen and not for four. Design's, round 8, and the argument is theirs: "four surfaces reached by four memorised URLs is four places for one to be forgotten, and the daily digest needs somewhere to link." Built as task 110, after the launch gate — 106's plain screens satisfy the gate, and a redesign is not a gate item.
D-36 A reported photo's verdict: Remove settles as flagged, Approve returns it to pending. Not hidden_pending_review, which promises a person will look when a person just did; and emphatically not pass, because a photo can reach the review queue having never been screened. Nothing — the path was unwired, and a reviewer's Remove silently did nothing to the photo. Fail open for the owner, closed for the public. pending puts it back in the retry sweep's path, which screens it within the hour. A removed value would read better than flagged and is what outfit_entries.moderation_status calls the same state; it was not taken because it is a schema change for a distinction review_queue.resolved_by and photo_screenings already record.
D-37 Development and deployment are separate sweeps (owner, 2026-09-25). All launch development lands first (docs/launch/development-plan.md, tasks 125–129); deployment is its own later sweep with its own plan (docs/launch/deployment-plan.md). docs/workflow.md's launch-gate checklist, which mixed code, dashboard and legal steps in one list. The production-readiness audit found both kinds of gap, and they have different owners, different lead times and different failure modes. One list hid which items a lane could close and which only the owner could.
D-38 Rollout is staged: the owner alone, then invited friends, then the public (owner, 2026-09-25). Each stage has a gate in the deployment plan; every development item is marked with the first stage it must be true for. The single "before public sign-ups" gate. Strava's capacity is 1, then 10 self-serve, then whatever review grants; the friends stage fits inside 10 while the review runs. Friends are also real users, so the gate before them is real rather than a formality.
D-39 Sign-up requires a valid invite code, and a "request access" form collects everyone else (owner, 2026-09-25). Codes are created and tracked in The Desk, and requests are listed there. Google sign-up needs a code too. Open email sign-up. The owner wants to see interest before opening the door. A code also bounds how many accounts a script can make, which Turnstile alone does not.
D-40 EU and UK users are supported, not geo-blocked (owner, 2026-09-25). Account deletion, data export and a GDPR-grade privacy policy are development work; processor agreements, representatives and the UK OSA risk assessment are deployment (legal) work. The audit's suggested option of US-only sign-ups at launch (§1.4). Strava's API Policy §7.3 already requires a GDPR-grade policy whatever the geography, and deletion and export are expected everywhere. Geo-blocking would save the legal half and little of the code.
D-41 Usernames replace display_name, and display_name is dropped in the same change (owner, 2026-09-25). This is a stated exception to resilience law 8 (expand→contract). Law 8's two-deploy rule for a drop. Expand→contract protects a running deploy from a schema it does not know. The app has never been deployed, so there is no running deploy and no data to carry. The exception holds only until the first production deploy: the change must merge before it, and after it law 8 applies to everything, this column included.
D-42 Email verification is in v1, sent with Cloudflare Email Sending on Workers Paid (owner, 2026-09-25). The owner authorises adding the send_email binding to wrangler.jsonc; task 126 adds it with a bindings-conformance assertion (law 10). "Verification is already planned" with no sender chosen; the forbidden-zone rule for this one binding. Password reset needs email anyway, and verification is the same plumbing. Cloudflare's sender keeps it on one bill and one binding. Its beta quota is a known limit; the deployment plan names a fallback.
D-43 Which emails exist at launch (owner, 2026-09-25). Transactional mail is always sent and has no switch: verify, password reset, email change, Strava broken or revoked, content removed, ban. The Strava "you ran, log it" reminder is on by default with an off switch. Nothing commercial at launch. Templates use React Email if it renders in workerd (task 126 checks first); otherwise typed functions producing HTML and text. R-1's text, which said the app "emails the user" when no email path existed. The reminder is the retention hook (decision D-14), and it is dead in-app for anyone not already in the app. Nothing commercial means no CAN-SPAM unsubscribe machinery yet.
D-44 PWA push is out of scope for launch (owner, 2026-09-25). — Post-MVP (Epic 203); email carries the reminder.
D-45 Photo size (register R-3): the browser downscales before upload, and the server caps decoded pixels (owner, 2026-09-25). The long edge is capped (about 2048px) inside the existing W3 canvas step; the server refuses anything over the decoded-pixel budget as the backstop. Register R-3's three open options. The browser step makes the common case small before it costs bandwidth; the server cap makes the OOM class impossible whatever a client sends.
D-46 Superseded by D-69. Public-entry photos are served through short-lived signed URLs, cacheable for their TTL (owner, 2026-09-25). That puts them in Cloudflare's cache, where the CSAM tool can see them, and a leaked link expires. Private-entry and closet photos stay private. Every photo keeps upload-time screening. The dashboard toggle and the NCMEC procedure belong to the deployment plan. Register R-71's framing, which assumed the dashboard toggle alone covered us; private headers on every photo. The audit (§1.10) showed the tool scans only what the cache serves, and ours served nothing it could see. Signing keeps the privacy property that private was buying.
D-47 Five lanes build the launch: 125 ops & platform, 126 accounts, 127 Strava & logging, 128 content & safety, 129 feed (owner, 2026-09-25). Shared rules in docs/tasks/125-129-launch-development.md. — Each lane owns a disjoint set of files; the two cross-lane edits (usernames, and delete primitives reused by account deletion) are sequenced in the shared packet.
D-48 A FormField's focus ring sits on the field border (outline 2px ink, offset −1px), as design round 26 item 16 draws it (owner, 2026-09-25). This overrides the Accessibility Contract §06's "offset 2px" for FormFields only; every other control keeps §06. Design is asked to amend the contract. Built by task 125 in ui/form.tsx and a11y.css. Accessibility Contract §06 as written, for one control class. A focused field shows one line instead of two, and the error band (not the ring) tells error from focus. The board is the better answer here, so the owner let it win rather than wait for the contract.
D-49 The Google button uses the official G, Google's colours and shape, and its label in Archivo, not Roboto (owner, 2026-09-25). No fourth font family. Task 126 checks Google's branding rule on the label font before building, and asks if Google requires Roboto. Round 26 item 13, which takes Google's spec whole, Roboto Medium included. The stack has three families by rule (CLAUDE.md); a fourth for one button is not worth it if Google permits otherwise.
D-50 An unverified account's entries save private, and verifying restores the runner's default (owner, 2026-09-25). Design's "queued shares" state (round 26 item 11) is not adopted, so docs/contracts.md's sharing model is unchanged: an entry is public or private, nothing else. Useful, report, email change and reset-by-email still wait for verification, as round 26 draws. Task 126, plus a note to design. Round 26 item 11's queue ("Shares when you confirm your email"). A third sharing state is contract-shaped: every feed read, the consensus aggregate and the OG card would have to learn it. Saving private gets the same protection (nothing unverified reaches another runner) with no new state.
D-51 Per-entry OG share cards are generated now, as round 26 item 22 draws them (owner, 2026-09-25): 1200×630, never the photo; private, deleted, banned or unverified entries get the default card. Rendering an image in a Worker needs a new library (workers-og/satori or similar); task 125 measures its bundle and WASM cost and reports it in its design doc before adopting it. The audit's "OG tags" (text meta only). A shared entry link that previews as a bare title wastes the one moment a stranger sees the product. The cost is a dependency, which the lane measures rather than assumes.
D-52 The privacy page uses contract values and design's link placements (owner, 2026-09-25). The 620 document measure and the existing lead step, not the board's 680 measure and 17/1.65. Linked from the signed-out footer, under the sign-up form, from Settings › About and from every email footer; not under the log-in form (design's call, accepted). The email binding (decision D-42) and the same-change display_name drop (decision D-41) were reaffirmed the same day. Task 126. Round 26 item 14's type values; register R-105's "both auth forms". Contracts outrank artboards (CLAUDE.md). On placement, design's reason stands: "logging in isn't consent to anything new."
D-53 The whole site is noindex until the public launch (owner, 2026-09-26). The default robots meta in modules/ops/og/site-head.ts says noindex on every page; at the stage 2 gate the landing page and other public marketing pages flip to index by their own route meta, and profiles and entries stay noindex (docs/launch/deployment-plan.md §11). Development plan open decision 1, left open by the launch plan. Runner names, kit and photos in a search index are a privacy call made once, at launch, not by default. A route's own robots meta replaces the root's in TanStack's head de-dupe, so task 129's per-route meta can override it.
D-54 Strava keeps only the athlete id and a refresh token, used solely to revoke (owner, 2026-09-26). strava_connections no longer stores the access token or expires_at; access_token, expires_at, status (its only other value was broken), refresh_failure_count and refresh_first_failed_at are dropped, as is strava_revocations.access_token, whose refresh_token becomes NOT NULL. The revoke drain has no access-token fallback. One migration, core 0025_drop_strava_access_token_and_refresh_tracking. An owner-approved destructive change, and a stated exception to resilience law 8 (expand→contract). Task 127. Law 8's two-deploy rule for a drop; register R-1's refresh-failure tracking and broken status. The same basis as D-41: nothing has been deployed, so there is no running deploy to protect and no legacy row to drain. The app never reads activity data, so a grant is only ever revoked — on disconnect, account deletion or deauthorization — and /oauth/revoke takes the refresh token, which does not expire until rotated. Holding less is the privacy floor. The exception holds only until the first production deploy.
D-55 The Google button's label stays in Archivo, not Google Sans (owner, 2026-09-27). Closes D-49's open check: Google's branding guidelines (2026-07-07) name Google Sans Medium but do not require it, and their "don't" list forbids a changed G, colours and ground, not another font. The G, colours and shape are Google's; the label is ours, in Archivo semibold. Task 126 PR 1 owner question 1; round 26 item 13's Roboto. A fourth family for one button is not worth it when Google does not require one.
D-56 An old handle is never reclaimable by another runner (owner, 2026-09-27). A handle a runner leaves stays in username_history for good; only its last owner can take it back. /@old answers "This runner changed their name." and never resolves to someone else. Task 126 PR 1 owner question 2 (the alternative was freeing a handle after N days). Otherwise a shared or bookmarked /@old link could start pointing at a stranger.
D-57 A reserved handle reads as "taken", with no suggestion (owner, 2026-09-27). The refusal is "@admin is taken. Try another." — the taken sentence without its "like @…" clause. No distinct reserved-name sentence. Task 126 PR 1 owner question 3. A distinct sentence would disclose the deny list one probe at a time.
D-58 "Public" means visible to signed-in runners on dialed.run, not to the open web (owner, 2026-09-26; clarifies D-53). A public entry or profile is one any signed-in runner can see; nothing is served to a signed-out visitor or a search index by being public. Task 126 ACC-13 (legal pages), and every place copy says "public". It is what noindex (D-53) and the signed-in feed already do; copy and the privacy policy must say the same thing.
D-59 A runner may delete an entry that is under review (owner, 2026-09-27). The delete goes through, and any open review row for the entry or its photos is settled as removed by the author (feed/retract.ts), so the reconcile sweep never re-queues it. Quarantined photo copies have no row and stay out of the runner's reach. Task 128's open question 1, built to its bracketed default. The runner owns their content; a pending report is not a hold on it. What a moderator needed to see is either gone from every feed already or is the quarantine copy, which the delete does not touch.
D-60 The server refuses a photo over 16 megapixels before decoding it (owner, 2026-09-27). photoRefusal reads the image header and refuses anything above 4096² (64 MiB decoded) with "Photo must be 16 megapixels or smaller." The browser downscales to a 2048px long edge first (D-45), so only a non-browser client meets the cap. Task 128's owner question on the SAF-2 server cap. A cap on decoded pixels, not bytes, is what bounds Worker memory; 16 MP leaves room for any real phone photo that bypasses the browser step.
D-61 The EMAIL binding is not "remote": true (owner, 2026-09-27). wrangler.jsonc binds send_email locally: a deployed Worker always sends for real, and dev and e2e get Miniflare's simulated sender. Task 126 PR 2a owner question (a). remote changes only local dev, where it needs a CLOUDFLARE_API_TOKEN — without one npm run dev and every CI e2e run refuse to start.
D-62 hello@dialed.run is a monitored address (owner, 2026-09-27). Every @dialed.run address forwards to the owner, so the From address on every email, and the appeal address on D4, reaches a person. Task 126 PR 2a owner question (b). A reply to an account email, or an appeal, must land somewhere a human reads (law 6).
D-63 An unconfirmed runner may reset their password, and a completed reset confirms the address (owner, 2026-09-27; amends round 26 #11, which held reset by email for confirmed addresses). authMail.resetPassword sends to any account; createAuth's onPasswordReset sets emailVerified. Task 126 PR 2a owner question (c). Spending the link proves the runner reads that inbox, which is all a confirm link proves; holding reset back left an unconfirmed runner who forgot their password with no way in.
D-64 Opening an unsubscribe link changes nothing; its landing's one Unsubscribe button does (owner, 2026-09-27; overrides round 26 #19's "opening it is the unsubscribe: no confirm button"). The GET reads the signed link and asks; the button POSTs. RFC 8058's List-Unsubscribe one-click POST, a mail client's own button, is unchanged. PR #119 review. Security scanners and corporate mail gateways prefetch every link in a message, so a GET that unsubscribes unsubscribes runners who never clicked.
D-65 Unsubscribe links are signed with their own secret, UNSUBSCRIBE_SECRET, not BETTER_AUTH_SECRET (owner, 2026-09-27). Read only through src/env; missing, no link is built, no optional email goes, every link is refused, and /api/health names it. The purpose label stays. PR #119 review. Rotating the auth secret no longer breaks every link in every inbox, and nothing auth signs shares a key with a link that is public by design.
D-66 An owed email is marked sent the moment its send lands; it carries no Message-ID of ours (owner, 2026-09-27; corrected in the second PR #119 review). The row records sent_at and the sender's message_id in one statement, and the drain deletes a sent row rather than sending it again. Additive migration 0031_add_outbox_sent_at_and_message_id. The first version also set Message-ID: <{dedupeKey}@dialed.run> so a retried send would collapse into the first at the recipient; that cannot work here. Cloudflare Email Service sets Message-ID itself and refuses one from the sender with E_HEADER_NOT_ALLOWED (developers.cloudflare.com/email-service/reference/headers), so every owed email would have thrown in production. An X- header is allowed but no mail client dedupes on one, so none is sent. PR #119 review (both rounds). Cloudflare Email Sending has no idempotency key or dedupe. The mark covers a Worker that dies between the send and the delete; nothing covers one that dies between the send and the mark, and that email goes twice.
D-67 An author always sees their own entries, including one under review (owner, 2026-09-27). An entry hidden pending review stays in its author's Following feed and on D, marked [UNDER REVIEW] on the card and on D (FEED-6). Nobody else sees it: bans, blocks and reports still hide it from everyone else. A removed entry and a private one stay out of Following, as before. PR #123 review of task 129: followingFeedStatement ran the author's own entries through publiclyVisibleEntry(viewerId), so the marker FEED-6 asked for on the card had no card to sit on. The author is the one person who needs telling why nobody else can see it; hiding it from them too reads as the entry vanishing.
D-68 Reporting a runner's profile hides them from the reporter's runner search and H straight away (owner, 2026-09-27), the way a reported entry is hidden from its reporter. Only for the reporter, in SQL (profileNotReportedBy in modules/safety), with an EXPLAIN test. PR #123 review of task 129 (FEED-7 covered a reported entry but not a reported profile). W1 promises the reporter an immediate hide whatever the review decides; a profile report owes the same.
D-69 No signed photo links; photos stay private behind sign-in (owner, 2026-09-28; supersedes D-46). SAF-7's signing code, its redirect and PHOTO_URL_SECRET are gone, and no photo is served with a public Cache-Control. Every photo keeps upload-time screening. CSAM coverage for public content is decided before the public launch (deployment plan §8). D-46's signed, cacheable URLs. Cloudflare's CSAM Scanning Tool compares only "content served … through the Cloudflare cache" (developers.cloudflare.com/cache/reference/csam-scanning/); a response a Worker builds never enters that cache, so the links bought no scanning, only a leak window.
D-70 A suspected-CSAM quarantine is silent and preserves everything for a year (owner, 2026-09-28). No bell row and no email to the uploader. The content is hidden at once: its rows are deleted from every table a read touches, in the same batch that copies them — the entry, items, tags, photo rows, the uploader and when R2 says the bytes were uploaded — to quarantined_content, a locked table only an admin reads (quarantinedContentFor); the bytes are copied to quarantine/…, served only to an admin. Kept 365 days (retain_until). The owner reports to NCMEC's CyberTipline by hand (deployment plan §8). The removal notice every other takedown sends; a quarantined flag on the live rows, which every read would have to remember to filter. Telling the uploader tips off the person a report is about; deleting the rows loses the context a CyberTipline report needs.
D-71 dialed.run is for runners 16 and over (owner, 2026-09-27; round 27 #12's "16 is an owner call"). The line under Au2 reads "dialed.run is for runners 16 and over." (round 27 #12). No date of birth is asked or stored: the line is the gate, and a report of an under-16 account is handled as any other takedown. Round 27 #12 left the age to the owner. Sixteen is GDPR article 8's default age of digital consent, so no parent's consent is needed, and the line gates without collecting a birth date. The line and the terms (the owner's text) ship with PR 2b-2's legal pages; this records the number.
D-72 The username rules, at claim time (owner, 2026-09-27). (1) A vendored open-source word list (LDNOOBW, CC BY 4.0, lib/profanity-words.ts) is checked instantly, as whole parts of a handle — split at _ and digits, digit disguises read back — never inside a word. (2) OpenAI's omni-moderation is then asked once about the handle as words; when it cannot answer (no key, timeout, error) the claim fails open to the list alone and the failure is reported. (3) Either refusal reads as taken with no suggestion, as a reserved name does (D-57). (4) A suggestion is the handle plus a running word (_runs, _miles), else a digit 2–9 — never the city. (5) No name is kept: Better Auth's user.name holds "" for every account, Google's profile name included, and migration 0038_blank_user_names clears the ones stored. (6) D-56 stands, and a deleted account's handle is never released: deletion (PR 2b-2) moves it to username_history like any other handle given up. The city-slug suggestion (round 26 #7, "like @x_pdx"); keeping Google's profile name in user.name. A suggestion built from the city told anyone who met a taken handle where its owner runs. A name nobody reads is personal data held for nothing. A list alone misses what is run together (bigshitrunner) and a model alone is down sometimes; together, the list is the floor.
D-73 An appeal of a ban or a removal carries no deadline, no timeframe and no "different moderator" (owner, 2026-09-29). The service has one operator. The ban email (account_closed) reads "Think we got it wrong? Reply to this email to appeal and we'll look again." (approved as worded in PR #129); the removal notice (content_removed) keeps "Think we got it wrong? Reply to this email."; D4, the banned runner's notice, reads "If you think this is wrong, write to desk@dialed.run. A person reads every message." — the board's "answers within a week" is dropped and the rest kept. Round 27 #15's "Reply within 30 days and a different moderator will look" (the Rulings board); Operator Screens D4's "answers within a week". With one operator, "a different moderator" is a promise nobody can keep, and a deadline or a response time is a commitment with nobody to cover it when that one person is away. Replying is the appeal; the promise is that a person reads it.
D-74 A garment that has runs can be deleted, and Y offers both actions (owner's round-3 ruling on PR #101, recorded 2026-09-29). Deleting a piece with runs opens round 26 #3's sheet: Retire it (the pink primary, recommended) and Delete it and its record (a hairline secondary), with what a delete costs listed above them. This narrows CLAUDE.md's "Retire, don't delete garments referenced by any entry" to "retire is the default"; CLAUDE.md itself is not changed by the lane that recorded this. CLAUDE.md's product rule read literally: no delete for a garment any entry references. The ruling was given without a written reason (PR #101's body, "Owner's rulings (round 3)", item 1); it is recorded so that a reader of CLAUDE.md's rule does not take the shipped Delete for a violation of it. Retire stays the recommended action, and the sheet says what a delete takes (the runs it comes off, its record in each band) before it runs.
D-75 F, the add-garment form, gets round 26's TYPE picker (owner, 2026-09-29). Not in PR #129 (task 128, PR 2b): a follow-up closet PR, register R-112. garmentSchema already carries an optional per-category type and wardrobe_items.type exists, so it is expected to be wiring; if it turns out to need a schema change, the owner is asked before it is made. Round 5's "type comes from the product, never a user", as register R-33 records it, for F. Round 26's F board draws the picker, and the owner chose to build it. The model question round 5 settled (a type per category, named for the pack's glyphs) is unchanged; only who may set it widens.
D-76 SAF-11's W3 lifetime goes to design round 28 (owner, 2026-09-29). W3 closes on the first onReady, which PhotoBlur calls as soon as the detector's blur is painted, so tap-to-blur and round 27's keyboard cells are on screen for a moment rather than until the runner is done. Keeping the step open needs a "use this photo" confirm, which is a design question: docs/design-deltas.md item 31 carries the ask, register R-113 the build. — The blur controls exist to be used, and a step that closes before they can be is a control nobody can reach; but a new confirm step is a new moment in the flow, which is design's to draw.
D-77 Three round 26 deltas ship as built and go to design round 28 (owner, 2026-09-29): GOES and SAME NAME in cold-text, STAYS and SAVED TO CLOSET in dialed-text, as round 26 draws them, against ink.css's "hue means verdict"; the PHOTO NOT ADDED kicker in FailureBand's grammar, with no hi-viz ground; and the rail's "Retired Mar 12" against the board's "RETIRED MAR 2026". All six of PR #129's undesigned surfaces and deltas are queued as docs/design-deltas.md item 38. — Each is the build's reading of a board against a contract or an existing pattern; design answers which wins, and nothing waits on it.
D-78 Reports a deleted runner filed are kept, with the reporter anonymised (owner, 2026-09-29; open question 3 of docs/designs/126-accounts.md). The purge rewrites reports.reporter_id to deleted:{report id}, so the report stays as the reported runner's record, names nobody, and one-per-reporter counts stay true. Already built in PR #130 (account/purge.ts); this records it. Deleting a leaving runner's reports with the rest of their rows. What was said about someone else's post is that runner's record too, and a moderation history with holes in it cannot be read.
D-79 The data export becomes an emailed ZIP before launch (owner, 2026-09-29), as round 27 #13 draws it: CSVs, original run files and photos, built by a queued job, stored in R2 and emailed as a 7-day link. A follow-up PR, not PR #130, which ships the packet's JSON download as the interim. Register R-116. The packet's and the development plan's JSON download as the finished ACC-10. The board is the product's answer; the JSON is honest and complete meanwhile, and the ZIP needs a queue type, an R2 prefix with a lifecycle rule and an email that are their own review.
D-80 Settings › About stays (owner, 2026-09-29): the group and its one row, "Privacy policy", ship as a placeholder and go to design round 28 (docs/design-deltas.md item 38). U1 draws no About group, so e2e/conformance/system-settings.conformance.spec.ts leaves that group and its row out of the comparison by name, and says why. U1's index as the complete list of Settings groups. The legal texts must be reachable from Settings (D-52), and U1 was drawn before they were.
D-81 The legal links show now, even while /privacy is a 404 (owner, 2026-09-29). The signed-out footer, Settings › About and every email footer link to /privacy, which answers X1 until the owner marks the text published (account/legal-markdown.ts' published mark). No change to the links. Hiding each link until its page exists. The links are the permanent structure; the page arriving behind them is one edit to the text's first lines, and a launch blocker in its own right (R-105).
D-82 A purged account's old /@handle says "This runner isn't here." (owner, 2026-09-29). Neutral: it reveals neither the deletion, nor a rename, nor a removal. A handle whose runner renamed and still has an account keeps D-56's "This runner changed their name."; a handle whose holder's profile is gone (the purge keeps the handle in username_history and deletes the profile) says the new sentence — both the handle they held at deletion and any they gave up before. Sign-in during a purge answers log-in's own "wrong email or password" for the same reason. D-56's "changed their name" for every handle given up. "Changed their name" would be false for a deleted account, and "deleted" would tell a stranger something about a person who has left.
D-83 The export ZIP keeps kit.csv and profile.csv (owner, 2026-09-30), beyond round 27 #13's three CSVs: each garment in each kit with its flag and note, and the account and settings as one row. Both are named in the README. A board delta, queued for design round 28 (docs/design-deltas.md item 39). Task 126 PR 2b-3. Round 27 #13's list of three CSVs. entries.csv cannot hold a kit's per-garment flags and notes without packing a list into a cell, and a copy of your data that leaves out your account is not a whole copy.
D-84 Settings › Export offers Download on the day a copy was asked for (owner, 2026-09-30), beside the emailed link; after that day the row offers a new copy, and the link works out its 7 days. So data_exports.link_token stays stored in plain text: the row reads it back to link where the email links, and it opens nothing without the owner's session. A board delta, queued for design round 28 (docs/design-deltas.md item 39). Task 126 PR 2b-3. Round 27 #13's email-only delivery; a token hashed at rest. The runner who asked is often still on the page. A hash would leave the row nothing to link with, and the session check, not the token, is the gate.
D-85 An R2 lifecycle rule deletes IMPORTS' exports/ objects 8 days after upload (owner, 2026-09-30), set at deploy time as a second net behind the hourly sweep. An owner deploy step (docs/deployment.md §2, docs/launch/deployment-plan.md §1); no code. Task 126 PR 2b-3. The design doc's "not needed" (owner question 3). The sweep deletes what rows name, but a build that dies between its put and its ready batch leaves a ZIP under a claim no row holds. Eight days is past the seven a link lives.
D-86 Data exports get their own queue, dialed-exports (owner, 2026-09-30): its own dead-letter queue dialed-exports-dlq, max_batch_size: 1, and max_retries: 3 as dialed-imports has. The account_export job moved off dialed-imports rather than living on both, which law 9 allows only because nothing has deployed: no message of the old shape can be in flight. The one wrangler.jsonc edit approved for PR #132; both queues are created before the first deploy that uses them (docs/deployment.md §3). Task 126 PR 2b-3. The design doc's "no new binding": a variant on dialed-imports. One ZIP build a delivery, so a long build never holds up a run file's parse, and a dead letter on its own DLQ is an export's and nothing else's.
D-87 The Desk has five destinations, as round 28 #14 draws them (owner, 2026-09-30): Today, Review, Access, Duplicates and Runners. Gave up (D6) is a section on Today, carrying its count, not a destination. Amends D-35's list of four. Build work for lane 125 (ops/components/DeskShell.tsx's rail, Today.tsx), not the round 28 import. D-35's "Today, Review, Duplicates and Runners". Access has its own queue, so it is a place to work; Gave up is a list to check, so it belongs on the page that is checked first. The build's rail already carries Access.
D-88 Removing suspected CSAM does not close the uploader's account (owner, 2026-09-30). Quarantine stays separate from closing an account and never closes one by itself, as D-70 and safety/quarantine.ts already behave; closing an account stays its own action on the runner. The Review row's second press says only what happens: round 28 #8's "close @n8's account" is not adopted. Build work for lane 128 (safety/components/ReviewQueue.tsx). Round 28 #8's confirm, "Remove it everywhere, close @n8's account and keep the evidence for the report?". Confirm copy must describe the action exactly; a quarantine that also banned would be a second, irreversible decision folded into a press about content. The board itself says the procedure is the owner's.
D-89 Reopening a closed account sends one email, "Your dialed.run account is open again." (owner, 2026-09-30), as round 28 #8 draws it. Transactional, with no switch. Amends D-43's list of launch emails. Build work for lanes 126 and 128 (a new email kind, sent through 125's outbox). D-43's list as the complete set of launch emails. A runner who was told their account was closed is owed being told it is open again; it is about their account, so it is transactional mail.
D-90 The author's under-review marker stays D-67's bracketed [UNDER REVIEW] (owner, 2026-09-30). Round 28 #6's plain hi-viz UNDER REVIEW tag, drawn without brackets, is not adopted. D-67's substance (the author always sees their own entry, and nobody else does) is unchanged. Round 28 #6's tag without brackets. D-67's wording stands until the owner changes it.
D-91 §4a failure kickers on hi-viz are not decided: sent back to design (owner, 2026-09-30). Round 28 #13 says every §4a kicker sits on the hi-viz ground; the Form Contract's failure band ("No fill, no yellow, values untouched") and docs/product.md §4 ("No fill. No yellow.") stay in force, and ui/FailureBand stays as built. The question is open queue item 40 in docs/design-deltas.md. Round 28 #13's "Every §4a kicker sits on the hi-viz ground. Put round 26's back." The ruling would change every failure band in the product, not one, against a contract that says why yellow is wrong there; design should say whether it means to amend the Form Contract or only notice bands.
D-92 D7's revoked row is not drawn with opacity (by precedence, 2026-09-30; no owner call). Round 28 #9 draws "reduced weight" as opacity: 0.55 over the row; T1 says "full strength, never opacity" and Accessibility 02 sets 4.5:1 for body text. The contract wins: the reduced weight is built from T1 roles (--quiet text, REVOKED in its state colour). Build work for lane 126 (account/components/DeskAccess.tsx). Round 28 #9's opacity: 0.55. Contracts outrank artboards (CLAUDE.md, "Design truth"); a muted mono line at 55% cannot hold 4.5:1.
D-93 No terms are accepted until they are published (owner, 2026-10-02). Until docs/legal/terms.md carries the published mark, "no published terms" is a state of its own: sign-up records no acceptance, nobody sees the terms prompt and no server function refuses anyone over the terms (account/terms-acceptance.ts currentTermsVersion). Once the mark lands, every account without an acceptance of that version meets the prompt once. Nobody ever accepts a text that was not public. Lane 126, PR #140. PR #140's first build, which recorded and asked about the draft's version 1 while the draft was unpublished. An acceptance of a text nobody could read is not consent, and the draft is not reviewed.
D-94 The terms prompt keeps an explicit Accept for a version bump (owner, 2026-10-02). terms.md's "If you keep using [dialed.run] after that, you accept the new terms" does not match: that is the owner's text, to fix in the pre-launch legal review, and is left as written; docs/legal/terms-sources.md records the mismatch (judgment call 13). Treating continued use as acceptance, as the draft's wording says. A recorded press is evidence of acceptance; continued use is not.
D-95 Get a copy is exempt from the terms gate, beside Delete account, Keep, Accept and sign-out (owner, 2026-10-02): a runner's copy of their data never waits on their accepting new terms. Settings › Account's two reads are exempt with it, as the page Delete account and Get a copy sit on. The export ZIP gains terms.csv, every version the runner accepted and when, named in the README. Lane 126, PR #140; a board delta (docs/design-deltas.md item 42). PR #140's gate, which refused Get a copy to a runner behind on the terms. Data portability is the runner's right, not a feature the terms can withhold; the acceptance history is their data too.
D-96 A stale tab's refused call goes straight to the terms prompt (owner, 2026-10-02). When a server function refuses with TERMS_NOT_ACCEPTED, the client opens /account/terms instead of a failure band, with no new copy, never saying the refused write was saved — from a form, a control or a loader (ui/terms-refusal.tsx). The root gate's has-handle memo is forgotten as it does, so the next in-app navigation asks the gate again. After Accept the runner returns where they were (from), or home. Lane 126, PR #140. "Our end failed. Nothing changed." for a refusal that is not a failure. The runner did nothing wrong and nothing broke; the one thing to do next is accept or leave.
D-97 A handle the re-ask flags gets Rename and Keep on its Desk review row, never Remove (owner, 2026-10-03). Rename is D8's moderator force-rename (round 27 #16), with its reasons list, and the runner then meets O0's "USERNAME CHANGED BY A MODERATOR" re-pick. Keep says the handle is fine and clears the flag (username_screen becomes clear). The row is marked from the profile's own username_screen = 'flagged', so a profile reported by people keeps the ordinary decision. Lane 126 builds it in lane 128's Desk code, through account's and safety's barrels (account/handle-review.ts, safety/components/ReviewQueue.tsx); PR #142. The re-ask's first build, whose flagged handle reached the queue with only Approve and Remove, and Remove writes nothing for a profile. A name is not a reason to close an account, and the force-rename already exists with the reasons and the runner's side drawn.
D-98 Change password's per-runner lockout (register R-124) is built before public, not in PR #142 (owner, 2026-10-03). Until then Change password answers Better Auth's own limiter with the band's "Too many tries. Wait a minute, then try again.", and round 28 #10's "That's 5 wrong tries" frame shows on Change email and Delete account only. Building it in PR B, or leaving it to after launch. It needs a server function in front of Better Auth's endpoint that verifies the password a second time, which is its own piece of work; the invite-only stages carry little risk, the public launch does.
D-99 Two hosts: dialed.run is the marketing site, app.dialed.run the app, api.dialed.run the read API (owner, 2026-10-03; round 30, Marketing M0). The marketing host becomes indexable when its SITE_INDEXABLE flag is turned on at public launch; the app host is never indexed, and its cookies are scoped to it. Amends D-53: the stage 2 flip becomes that flag on the marketing host, and profiles and entries stay out of the index. The deployment sweep moves BETTER_AUTH_URL, the Google OAuth redirect URIs, the Strava callback domain, cookies, every email link and the CSP to the new hosts. D-53's single site, where route meta flips the landing page to index at stage 2; the deployment plan's one Worker on dialed.run. An indexable site and a never-indexed app are easiest to keep apart as two hosts: one robots rule per host, and no app cookie ever reaches marketing pages.
D-100 The legal pages move to the marketing host once the marketing site exists, sourced from this repo's docs/legal (owner, 2026-10-03; round 30, Marketing M0). Until then the app keeps serving /privacy, /terms, /copyright and /open-source. Amends D-81 and D-93 only as to host: the links, the published mark and "no terms are accepted until published" are unchanged. Round 30 M0's static legal pages, read as a change to D-81's links and D-93's published mark. The legal text has one source either way; only where it is served changes, and only when there is somewhere to serve it.
D-101 Guides and reports on the open web are anonymous totals, not entries (owner, 2026-10-03; round 30, Marketing M1–M3 and M8). D-58 still governs entries and profiles: nothing of an entry is ever public. Whether published totals count closet pieces (D-29) is still open (docs/design-deltas.md, round 30). Reading D-58 as forbidding any figure derived from shared entries from reaching a signed-out visitor. A total of five or more runners names nobody; D-58 is about what a person shared, and a total shows no one's run.
D-102 The terms prompt says where Accept returns, and that typed values were not kept (owner, 2026-10-03; round 30 #4b): "After you accept, you'll go back to Log a run. What you typed wasn't kept.", shown only when a refused save brought the runner there. Amends D-96, whose point was no failure band, not no words. Lane 126 (account/components/TermsPrompt.tsx). D-96's "with no new copy". Losing what was typed is no surprise if the page says so before it happens.
D-103 API tokens keep reading while a runner is behind on the terms, and Revoke stays allowed (owner, 2026-10-03; round 30 #6), consistent with D-95's export exemption. Creating a token waits on Accept. Task 130's design doc changes to match (it refused with 403 terms_not_accepted). The read API design doc's terms gate on every token request. Reading your own data is portability, which D-95 already says never waits on new terms; revoking is a safety action.
D-104 Creating an API token sends a security email, transactional with no opt-out (owner, 2026-10-03; round 30 #6): "A token called {name} was created on your account.", with its scope and a "This wasn't me" link that revokes it. Amends D-43's list of emails. Built with the read API (task 130). D-43's list as the complete set of emails. A credential that can read private runs should never appear on an account without its owner hearing about it.
D-105 The marketing site is a separate public repo built with Astro, dialed.run-site, deployed to the same Cloudflare account with a scoped deploy token (owner, 2026-10-03; round 30, Marketing M0–M8). Guide data reaches it as a nightly aggregate JSON that an app cron writes to R2, holding only bands with 5 or more runners. Changelog entries are drafted by Claude from merged PRs and approved by the owner. A public-launch item. Building the marketing pages inside the app worker. A static site needs none of the app's bindings, and a separate repo keeps it out of the app's gates and bundle; the R2 handoff means the site never reads D1.
D-106 No kit line on the Strava activity (owner, 2026-10-03; Integration Opportunities 04, not adopted). D-14 and D-54 stand: no activity data, and the refresh token is used solely to revoke. The Integration Opportunities board's opt-in "41°F damp · Dialed · kit via dialed.run" line. It needs Strava's write scope and would use the grant for more than revoking.
D-107 Link previews use the default card only (owner, 2026-10-03; round 27 #7, restated by round 30 #5). No per-entry OG card is generated, and a signed-out entry link previews as the default card. Supersedes D-51. D-51's per-entry OG cards. The owner's earlier "generic card only" call: a preview must never carry entry data, which is D-58's rule for link previews too.
D-108 Published anonymous totals count shared and private runs, by the middle path (owner, 2026-10-03, with design's round 32 conditions). A: totals read only the feels-like band, sky, month, garment type and catalogue model, the verdict, and region if set, and nothing else. B: every published figure needs 5 or more distinct runners, or 20 for a brand. C: confirmed accounts only, without removed or quarantined entries, banned or leaving accounts, or manual-temperature runs. D: Settings › Sharing › "Count my runs in anonymous totals" is on by default; off leaves the guides at the next nightly build, and published reports stay as they are. E: disclosed under A3's PRIVATE state, in Settings and in Au2's legal line. Closes the D-29 question: closet pieces count only when worn on a counted run, and in-app social proof stays shared-only. If A or B can't be met at launch, ship shared-only. The Terms and Privacy wording is the owner's legal text. Shared-only totals, the D-58 reading round 30 assumed; whether reports count closet pieces (#144). A guide built only from shared runs overstates "dialed". Deciding before launch costs one Terms version, and after launch a second acceptance prompt.
D-109 Visibility is an audience, not a flag (owner, 2026-10-03: groups are a Day 2 item, with a minor change before launch to make room; design round 33, design/docs/product.md "Pre-launch: audience model"). An entry's sharing choice is `audience: 'private' 'groups' 'runners'in place of theis_publicboolean, and the runner's default is an audience in place ofshare_default. Every audience check reads it, never a boolean: the feeds, the entry page, the read API's visibility, social proof, the export, and the anonymous totals. Launch uses only privateandrunners, and the copy does not change (SHARED is runners, PRIVATE is private). Groups themselves are parked, not scheduled: design/Day 2 Groups.dc.html is direction only, and its seven open questions are undecided. User-facing files write the UI's words (shared, private, later groups), and runners` stays internal. Group membership is stored when groups ship, as a join table. The build is R-129.
D-110 Original run files are kept for as long as the run exists (owner, 2026-10-04, reviewing design 130). This replaces the 30-day expiry on dialed-imports. An uploaded .fit, .gpx or .tcx is deleted when its run is deleted or the account is, by the deletion paths, not by a bucket rule. The bucket's lifecycle rule narrows to the 8-day exports/ prefix, which is an owner dashboard step. The privacy policy's "Imported run files: 30 days" changes to match, and the owner reviews it as legal text. A failed import's file is deleted 30 days after the import failed (owner, 2026-10-04): it never became a run, so nothing else would delete it, and 30 days leaves a window to debug a parser bug. Why: expiry cannot be undone, and keeping the files is what lets new fields be re-parsed from old runs and leaves routes and maps possible later. The build is R-131.
D-111 Runs record moving time and elevation gain, and TCX reads every lap (owner, 2026-10-04, reviewing design 130). Two nullable columns are added to runs, moving_s and elevation_gain_m. moving_s is FIT's totalMovingTime when the watch writes it, else totalTimerTime; for TCX it is the laps' summed TotalTimeSeconds; for GPX it is the time between points with the pause gaps left out. elevation_gain_m is FIT's totalAscent, or a smoothed sum of point altitudes for TCX and GPX. Treadmill and manual runs leave both null. durationS stays elapsed time, which keeps running through a pause. The read API returns durationS, movingS and elevation in metres and feet, and computes pace from moving time when there is one. Runs whose file is still held are re-parsed to fill both columns. Nothing on screen changes. The build is R-130.
D-112 M5 carries Au5's optional note (owner, 2026-10-04, asked by the marketing-site agent). The marketing site's invite page gets the same optional note field Au5 has, because the Desk's access review shows the note to whoever decides on a code. At most 140 characters (ACCESS_NOTE_MAX), with a counter from 120. It is posted as note to the public access-request endpoint, which answers invalid_note over 140 with the site's copy, "Keep the note under 140 characters." The site has built it to match Au5; design-deltas item 48 asks design to confirm its placement.
D-113 Writes that others can see, or that trust the address, need a confirmed email (owner, 2026-10-04, approving design 133). One gate, auth's verifiedUserId(), refuses with EMAIL_UNCONFIRMED; an architecture test classifies every server function and route, both ways. The client opens "Confirm your email first" from one root handler. Q1: an unconfirmed runner's garment keeps its typed brand and name with no product_id; confirming owes an outbox job that links them. Q2: handles may be claimed, but unconfirmed runners are hidden from search and H. Q3: follow needs it, unfollow does not. Q4: private entries, verdicts, captions and photos do not (D-50's clamp stays); whatever other runners see or that combines runners reads only confirmed accounts' shared entries. Q5: the Desk needs it. Q6: export and deletion do not (D-95). Q7: block does not, report does. Q8: email change does. Design 133 §Decisions has the detail.
D-114 dialed.run is for runners 18 and over, and sign-up asks the date of birth (owner, 2026-10-09; design 134). Replaces D-71's 16 and its "no date of birth is asked or stored". Au2 asks the date on both ways in (email and Google); the server checks it, keeps neither the date nor the age, and a refusal sets a day-long cookie (dialed_age_refused) so a different date is not tried straight away. D-71's line under Au2 says 18. An under-18 account found later is closed and its data deleted. Accounts made before this shipped are not asked: pre-launch, every one is an invitee the owner knows. Why: the launch legal research (2026-10-09) found that admitting 16- and 17-year-olds brings Nebraska's Parental Rights in Social Media Act (LB383) into reach — its parental-supervision duties (§28(4)) have no size threshold and are not enjoined — and Mississippi's and Tennessee's similar laws, and every minors law found draws the line at 18. The owner has no reason for under-18s to use the app.
D-115 Sign-up stays open to the EU, EEA, UK and Switzerland; no EU or UK representative is appointed (owner, 2026-10-09). An accepted risk, not an oversight. GDPR probably reaches the app through article 3(2)(b) (logging runs and locations is monitoring behaviour), which formally needs an article 27 representative; the UK's Online Safety Act expects a written illegal-content and children's-access risk assessment. Neither is done today. Why: no enforcement against a tiny non-targeting non-EU app was found, Strava's API terms already require a GDPR-grade privacy policy whatever the audience, account deletion (GDPR art. 17) and the export (arts. 15 and 20) exist, and blocking would turn away the few European runners for little gain. Revisit if European sign-ups become a meaningful share, if a regulator or user asks, or if the app starts making money. The cheapest exit is a Cloudflare rule on the sign-up path. Triggers (owner, 2026-10-18; legal review rev. 2): appoint EU and UK representatives at the first EU/UK complaint, rights dispute or regulator contact; at 250 EU/UK accounts or 20% of the base; at any breach touching an EU/UK user; when the Call makes any health-adjacent inference; or at the first EU/UK revenue. The geo-block rule is written and left disabled as the fallback.
D-116 An imported run file is deleted once it has been read (owner, 2026-10-09; design 135). Replaces D-110's "kept for as long as the run". The batch that marks an import done or duplicate owes the file's deletion, due at once (import_file_expire), and the daily drain pays it. A failed import's file still waits 30 days. The run keeps what was read, with its start point rounded to two decimals as before (R-110). What this gives up: re-parsing old runs for a new field (D-111's kind of change) and drawing routes from stored files (R-132) — both now need the runner to upload the file again, which is R-146. Why: a full GPS track usually starts and ends at a front door, so it is the one precise location the app still held. The launch legal research (2026-10-09) found that storing precise geolocation (within 1,750 ft) brings Connecticut's privacy act into reach from its first Connecticut user (sensitive data, no size floor since 2026-07-01), and Washington's health-data law and the FTC's unfairness theory press on the same data. Rounded start points are outside that radius; the files were not. Nothing is live, so there is nothing to clean up.
D-117 Intimate images shared without consent have their own removal path (owner, 2026-10-10; design 136, R-145). In the app: a report reason, intimate, that hides its kit or photo from everyone on the first report rather than the third, and emails every operator when its removal is due, 48 hours on. A person still decides. Out of the app, for someone with no account: a public page and form (/takedown, PR B) whose request queues with an alert but hides nothing until an operator looks, because a form anyone on the internet can fill should not be able to hide any photo. Amended 2026-10-10 (round 35 #58e, owner): design drew the request as hiding at once, "as a W1 report of the reason does", and the owner took that over the queue-only answer: a request whose link is one of ours now hides it at once, and a Decline gives it back unless a review row or another pending request still holds it. A link that is not ours hides nothing. The cost accepted: anyone, with no account, can hide one kit or photo until an operator looks, within 48 hours; Turnstile and the per-address hourly limit are what stand in front of that. Requests are kept until the owner deletes them. Copies: the Review row lists the same runner's other photos; there is no photo hashing, and blur re-encodes every upload, so an exact hash would rarely match. Why: the TAKE IT DOWN Act (enforced since 2026-05-19, no size floor) requires a clear and conspicuous request path, removal of a valid request within 48 hours, and reasonable efforts on known identical copies. The cost the owner accepted: one confirmed account can now hide a kit or photo until a person reviews it, where reports.ts had reasoned that three were needed to stop a takedown on demand.
D-118 /takedown has abuse limits (owner, 2026-10-10; design 137). Since D-117's amendment a request whose link is ours hides it at once, and anyone can file one. Requesters are matched by a normalised email (validator's normalizeEmail, plus any +tag) and by an HMAC of the visitor's IP under a new secret, TAKEDOWN_SECRET (email alone until it is set). Cap 2: a requester with two open requests files a third that waits for a confirm link before it hides or alerts. Valve 10/hour: past ten hides in an hour site-wide, new requests hide nothing and every operator gets one alert per hour. Blocked requesters (Review's Block and decline all) get the ordinary receipt and nothing else. Hide-at-once with no limits; a refusal for blocked requesters (rejected: it tells an abuser they are blocked). A form anyone can fill must not be able to take the site's photos down faster than a person can put them back.
D-119 Published anonymous totals go out under CC BY 4.0, on legitimate interests (owner, 2026-10-11, answering PR #182). The guides' and reports' totals (D-101, D-108) are published as data anyone may reuse, under the Creative Commons Attribution 4.0 licence. They are a marketing item, not sold. Where GDPR or UK GDPR applies, the legal basis is legitimate interests, with Settings › Privacy's switch as the way to object; the balancing note is docs/legal/aggregates-lia.md. The Terms licence and the privacy policy's "Anonymous totals" section say both. Consent as the basis; totals published without a reuse licence. Consent needs an off-by-default setting, which brings back the self-selection bias D-108 counts private runs to avoid. An open licence suits totals that hold nothing about any one runner.
D-120 Totals group by climate zone and usual conditions, not a self-declared region (owner, 2026-10-11; amends D-108 A). "Region if set" is replaced by a broad climate grouping derived from the run's location (a Köppen-Geiger zone, never anything more precise), and/or a grouping of runners by the conditions they usually run in, derived from their own counted runs. There is no self-declared region and no Region setting. A guide for a state is made on the marketing site by mapping the state to the climate zones it contains ("runners in climates like Minnesota's"), using no runner data. The zone is looked up when the guide is built, from the run's rounded coordinates, with nothing new stored (R-134). D-108 A's "region if set"; round 30 data requirement (d)'s Settings › Region. Climate varies too much within a state, and what region was meant to add is acclimatization, which a zone or a conditions baseline measures better. A zone is coarser than a state, so the privacy balance improves too.
D-121 Operators may access private content for named purposes; the privacy policy discloses it (owner, 2026-10-12). "Private" means not shown to other runners or the public. The people who run dialed.run may see any content, private kits included, when they need to for: running and securing the service; reports, removal requests and moderation, including checking a reported runner's other photos for copies (D-117, PR #188); legal obligations and requests; and support the runner asks for. Nothing else, and only those people. The privacy policy's "Our access to your content" says so, and its private-kit statements point to it. A narrow "checking for copies" exception, and "never shown to anyone else". A promise that nobody sees private content is one the operators cannot keep, and naming purposes is what GDPR purpose limitation and the app-store privacy labels ask for.
D-122 No counsel before public launch; legal and privacy calls are the owner's best judgment (owner, 2026-10-17). dialed.run is a side project and a lawyer is not justified yet. The privacy policy's [OWNER: for counsel] notes are answered in the text instead, and its DRAFT line ("not reviewed by a lawyer") stays, because it stays true. The floor that does not move: the policy says only what the code and the deploy steps make true, and privacy-policy-sources.md maps every claim. Risks are ranked by what would realistically bite: an untrue policy first (FTC Act §5, state consumer law, and the Google OAuth and Strava reviews that read it); then data-subject requests, which export and deletion already answer; then GDPR transfer mechanics, covered by each processor's DPF certification or SCCs. Amends D-115's "counsel" wording, not its accepted risk.
D-123 D1 read replication is on for both databases from launch (owner, 2026-10-17), so the owner and friends test the state the app will run in; Miniflare has no replicas, so production is the only place lag exists. The primaries stay enam; replicas go to every D1 region, and the privacy policy says so ("read-only copies … in Europe, Asia-Pacific and Oceania"; R2 is not replicated). Consistency: one D1 session per request, started from the runner's latest bookmark, so a runner never reads older than their own last write on any device, including writes queues and crons make for them. Stale views of other runners' changes are accepted (blocks, private, hides, takedowns, bans): a replica is always an earlier consistent state, so nothing posted after a block leaks past it, and the window is the replica lag (tens of ms within NA and EU; seconds at the tail to Asia; no bound during an incident). Not accepted: a stale check that authorizes a write. Auth gates, and any check that decides a mutation (block, audience, ban, deletion state, handle availability), read the primary or run inside the write. Design: docs/designs/143-d1-read-replication.md.
D-124 Void (VoidZero, 2026-10-15) is not adopted; watch it (owner, 2026-10-17). Framework mode wraps the same stack (@cloudflare/vite-plugin, Drizzle, Better Auth) and gains little. Native mode replaces TanStack Start but falls short of rules this codebase enforces: Link takes href: string, useParams<T>() is an unchecked cast, Void-managed queues get no dead-letter queue, there is no documented CSP nonce or workerd test story, and it reached 1.0 two days before review. Revisit when it has typed links and params, DLQ configuration and a workerd test path. Register R-155.
D-125 The friends launch waits for Big Long Run Labs LLC; Terms and Privacy v1 name the LLC as operator (owner, 2026-10-18; plan docs/launch/legal-readiness-plan.md). There is no own-name version. The LLC is a Nebraska single-member LLC; filed 2026-10-18, approval pending. The drafts' "run by one person, not a company" and the review's own-name Variant 1. One version, one acceptance: naming the LLC from v1 avoids a v2 and a re-accept prompt weeks after friends join.
D-126 The published address is the LLC's registered (home) address, on the terms, privacy policy, copyright page, DMCA directory and email footers. The DMCA agent's phone is a free Google Voice number (owner, 2026-10-18). Rev. 2's virtual mailbox (B2). The address is already public through the Secretary of State record, so a mailbox buys friction, not privacy. Real privacy needs a commercial registered agent plus a mailbox (R-157).
D-127 No invited friend is in the EU, EEA, UK or Switzerland (owner, 2026-10-18). The GDPR rights, legal-bases and transfers text still ships in Terms and Privacy v1. The Sentry DPA, the UK OSA assessments and the DSA reporter notice (D-132) are public-stage items. — Adding GDPR text after friends accept would be a second acceptance prompt (D-93); the operational EU items have no EU user to serve until public.
D-128 Strava capacity is 10 athletes and the owner is one of them: at most 9 friends connect Strava (owner, 2026-10-18). Request the 9,999 level before public. Strava refuses the next connection itself (STR-6), so no code cap is added. The friends-stage line "no more than 10 friends connect Strava" (deployment plan, Stage 1). The owner's own connection counts toward the 10 (legal review rev. 2, §3.3).
D-129 CSAM hash matching is PhotoDNA, with Thorn's Safer as the fallback (owner, 2026-10-18). Photos stay private behind sign-in. Resolves D-69's open item. The public cached bucket for Cloudflare's CSAM Scanning Tool (deployment plan §8, option 1). A public bucket would break the privacy policy's "never served from a public cache" and needs a binding and zone change; PhotoDNA matches at upload beside the classifier.
D-130 A copyright removal hides the material for 30 days, then deletes it; a valid counter-notice restores it 10–14 business days after it arrives (owner, 2026-10-18; plan C5). Removal deletes at once. 17 U.S.C. §512(g)(2)(C) safe harbour requires restoration after a valid counter-notice, which deletion makes impossible.
D-131 Takedown requests, reports and moderation-decision records are kept for 3 years after the decision, then swept (owner, 2026-10-18; plan C7). Quarantined CSAM keeps its own clock: one year after the NCMEC report (18 U.S.C. §2258A(h) as amended by the REPORT Act). Kept until deleted (indefinitely); quarantine's year counted from quarantine. Three years covers the DMCA limitations period, repeat-infringer and ban-evasion history and appeal disputes, and answers GDPR storage limitation.
D-132 A reporter is told the outcome of their report: an in-app notification, "We removed it" or "We reviewed it and left it up" (owner, 2026-10-18; plan C8). Public stage. No reporter notice. EU DSA Arts. 16(5) and 17 while EU sign-up is open (D-115).
D-133 Published anonymous totals add secondary suppression, round counts to the nearest 5, and rebuild guides at most monthly (owner, 2026-10-18). Amends D-108 (D): an opt-out takes effect at the next guide rebuild, within a month, not the next nightly build. D-108 (B)'s thresholds alone, and (D)'s nightly timing. Thresholds do not stop differencing two published figures (or two nights' figures after an opt-out) to isolate one runner (legal review rev. 2, 5b).
D-134 Disconnecting Strava inside the app sends the strava_disconnected confirmation email, as a Strava-side deauthorize already does; exactly one per disconnect (owner, 2026-10-18; plan C3). Account deletion does not send it. Email only on the webhook path. Consistent confirmation, and Strava's API Policy §2.5 deletion confirmation whatever its exact wording turns out to be.
D-135 The terms' content licence lets us "operate and develop" the service and display, reformat and modify links in content, including outbound product links; the spam rule reads "spam, advertising or promotion we have not allowed" (owner, 2026-10-18). The draft's "run" purpose line and "spam or ads". Written into v1 so later changes to how links are handled are not a material change forcing re-acceptance (D-93).
D-136 Disputes: Nebraska law; venue the District Court of Sarpy County or the U.S. District Court for the District of Nebraska; a US$100 liability cap; no arbitration, class waiver or user indemnity; a 30-day write-to-us-first step and a small-claims carve-out (owner, 2026-10-18). The drafts' open brackets. The owner lives in Sarpy County. Arbitration buys little for a free app, invites mass-arbitration fees and does not bind EU/UK consumers.
D-137 No separate subprocessor or credits pages: the privacy policy's processor table gains company, role, location and safeguard columns, and the credits (Visual Crossing, MediaPipe, the profanity list, the climate-zone dataset) render on the app's /open-source page, which the marketing site already links to (owner, 2026-10-18; plan C9). Rev. 2's /subprocessors and /credits. Derive, don't mirror: one table, one notices file. The marketing site's link to app /open-source is dead today.
D-138 A USPTO knockout search for DIALED.RUN before public, and no filing yet; no insurance now (owner, 2026-10-18). Insurance is R-158. Rev. 2's B14 (buy if ≤ ~$1,000/yr). DIALED alone is crowded (Trek, Ad Hoc Labs). A four-figure premium is not justified for a free side project before revenue.
D-139 Destructive schema changes are pre-approved until the first production deploy (owner, 2026-10-18). Drop or rename in one PR and say so in its description; numbering (law 11) still applies. Once friends are live, CLAUDE.md's stop-and-ask rule returns. CLAUDE.md's stop-and-ask for destructive changes, for the pre-deploy period only. Nothing is deployed, so expand→contract protects no running code or data.
D-140 UK sign-up stays open, and the UK Online Safety Act records are written before public (owner, 2026-10-18; legal review rev. 3 item 12): an illegal-content risk assessment across Ofcom's priority harms, a children's access assessment recorded as "likely to be accessed by children" (self-declared age is not age assurance), and a children's risk assessment with proportionate children's safety measures, each on Ofcom's record-keeping template and reviewed at least yearly and before any significant change. Any Ofcom information request is answered within its deadline. Rev. 2's single short assessment concluding "not likely to be accessed by children"; the "no UK links yet" option. Ofcom's Children's Access Assessments Guidance (3.9, 4.2, 4.11) does not allow that conclusion on an 18+ term and a date-of-birth prompt. Every Ofcom penalty so far began with an unanswered letter or no written assessment.
D-141 New sign-ups and access requests from Mississippi and Tennessee are refused now, by Cloudflare's region code at every way to create an account (email, Google, invite redemption); existing accounts and sign-in are unaffected. Access requests from those states are accepted and tagged with the region, so the Desk shows the demand, and are never sent an invite while the block stands (owner, 2026-10-18, after a second research round: a forced-private "solo" account is still an account on a covered service under both laws, so it is no way out) (owner, 2026-10-18; plan C13). The block lifts for a state only on an order enjoining its law as to non-parties; the dockets are checked monthly (R-159). Stripe Identity (US$1.50 per completed check) is the recorded way back in if demand appears. Rev. 2's "unless confirmed enjoined"; deferring to public. Miss. HB 1126 (5th Cir. stay, SCOTUS 25A97) and Tenn. §47-18-5703 are in force against every provider with no size threshold and require verifying every account creator's age (legal review rev. 3 item 1). Bluesky and Dreamwidth blocked Mississippi for the same reason.
D-142 A perceptual hash of every uploaded photo, computed from the original bytes before re-encoding, is stored and deleted with the photo, and used only to find copies of a reported or removed image (owner, 2026-10-18; plan C16). Before public. The privacy policy's photo-fingerprint bullet, retention and legal-bases rows ship with it. Manual search only (the uploader's other photos and supplied links). The TAKE IT DOWN Act requires reasonable efforts to remove known identical copies; the FTC points to hashing (legal review rev. 3 item 4).
D-143 Strava connect shows a consent disclosure before the OAuth redirect, with the owner-approved copy (what, how, withdraw, delete, confirmation) (owner, 2026-10-18; plan C14). Before friends. A privacy-policy link only. Strava API Policy §2.1 (1 June 2026) requires consent that discloses those five things before accessing data.
D-144 Weather attribution reads "Weather data provided by Visual Crossing", linked to visualcrossing.com, beside conditions on every view, share card and email that shows them and under every published guide; the terms and each guide carry Visual Crossing's pass-through accuracy disclaimer (owner, 2026-10-18; plan C15). The unlinked "Weather by Visual Crossing" (src/ui/WeatherAttribution.tsx). Visual Crossing's licence (editions footnote; Terms §19) fixes the phrase, requires the link and a materially equivalent disclaimer (legal review rev. 3 item 7).
D-145 DSA statements of reasons cover automatic hides too: an author gets a short in-app notice when a report hides their content ("hidden while a person reviews it"), and the existing removal and closure messages carry DSA Art. 17(3)'s elements (what, why, automation, rule, redress) (owner, 2026-10-18; plan C8, with the reporter notice of D-132). Before public. Suspected CSAM stays silent. Notices on removal and closure only. DSA Art. 17 has no micro-enterprise exemption and covers any restriction of visibility (legal review rev. 3 item 11).
D-146 A court action is a litigation hold: once the Desk records that a copyright notice's sender has filed a court action, the material stays hidden and is not purged at day 30. It stays held until the case ends and the Desk releases it with "Case over: delete" (a strike) or "Case over: restore" (the claimant lost, so no strike) (owner, 2026-10-23; plan C5, design 149). The 30-day purge goes ahead after a court action. Deleting the subject of a pending suit risks a spoliation claim.
D-147 A photo removed as an intimate image shared without consent keeps its fingerprint for 3 years, so a re-upload is caught (owner, 2026-10-24; amends D-142, design 152). Only removals with the reason intimate. A new kit photo within the match distance uploads but is held hidden_pending_review with one Review row: never refused, never removed automatically. Swept with D-131's records. Garment photos are not checked. The algorithm stays dHash; outside hash lists (StopNCII) are not joined. D-142's "deleted with the photo" for these images. Someone who has posted an intimate image here once is a real re-upload risk, and the FTC's guidance suggests hashing to prevent reappearance. Holding for review rather than refusing keeps a false match from silently blocking a runner, at near-zero review load.
D-148 The UK OSA accountable person, the evidence store, the sole admin and the CSAM closure email (owner, 2026-10-24). (1) The owner is the person accountable for the UK OSA illegal-content and children's safety duties; the public records say "the owner" and the signed copy names them. (2) What the compliance procedures keep outside the app (an NCMEC report's captured context for 18 U.S.C. §2258A(h)'s preservation, takedown notes, and the GDPR and UK GDPR Art. 33(5) breach register) goes in a private R2 bucket, dialed-evidence, with bucket locks: ncmec/ and takedown/ 400 days from upload, register/ indefinitely. It has no Worker binding; the owner reaches it with an R2 API token scoped to that bucket, and may encrypt objects with age first. Objects are deleted by hand when the record they back is swept. (3) There is no backup admin: an accepted risk, and TAKE IT DOWN's 48-hour removal is the obligation most at risk while the only admin is away. (4) Closing an account for suspected CSAM sends a generic email, "We closed your account for breaking the community rules." and the can't-log-in line, naming no reason and no rule (built in the C8 follow-up). The compliance drafts' [OWNER: …] markers for each; the operator-typed ban reason for suspected CSAM. Ofcom's Codes expect one named accountable person. R2 is a provider the service already uses, so no new processor; a lock refuses deletion and overwrite for its period, even by a lifecycle rule, so a mistake or a leaked token cannot cut §2258A's year short, and with no binding no app bug reaches it. 400 days covers the year from the report, since the context is captured before the report is filed. An account admin can still remove a lock rule, so it protects against accidents, not against the account. A second admin would mean a trusted person with access to reported images; at this size the owner accepts the missed-clock risk. A generic closure line does not tip off a suspected uploader.
D-149 A run imported through Garmin's API is the runner's own run, like a file they uploaded: it may be shared to the feed, with Garmin credited on every surface that shows it (G1). Phase 1 is the owner's alone, through their Personal API key (plan 160, owner-approved 2026-10-24). D-14's rule for Strava, which keeps activity data out of the app entirely. Strava's terms forbid showing its data to other users. The Personal API's terms bind only the key (personal to its holder, for their own tools); the runs it fetches are the runner's own watch's.
D-150 Bracket notation is drawn, not said. Bracketed marks its [ and ] aria-hidden, so the accessible name is the value alone: [Nothing in here yet] is announced "Nothing in here yet", [3–16°] "3–16°" (owner, 2026-10-31; plan 161). A tag that needs a fuller spoken sentence keeps UnderReview's pattern: the bracketed tag hidden whole, the sentence in sr-only. The Accessibility Contract, which says nothing on it, and the brackets in every accessible name since task 112. JAWS says "left bracket … right bracket" around every value by default and NVDA does at "most"; VoiceOver only pauses. A survey of every e2e state (plan 161) found no bracket touching a word, so hiding them joins no words. The cost is braille, which loses the glyphs, and the pauses VoiceOver made, which matter only where tags share a line with prose: the Desk's review rows, which get spoken separators of their own.
D-151 A runner can put a kit, and separately their profile, on the open web (owner, 2026-10-25; proposal 163). D-109's audience gains world, chosen per kit; runners keeps meaning signed-in runners, so nothing shared today becomes public. A public profile is its own Settings switch, off by default, with a second switch for the closet as a rotation. world can be a runner's default audience only once their profile is public. A public kit shows the kit, verdict, feels-like and sky (with D-144's credit), distance, moving time, pace, elevation gain, caption, handle and the run's day; never the time of day, location, route, per-item flags or notes. A choice of coarser dates can come later if runners ask. noindex on people's pages (D-53) stands. Supersedes D-58 in part: "public" can now mean the open web, for what a runner opts in. D-58's "public means signed-in runners, never the open web" for opted-in kits and profiles. A link that ends at a sign-in or invite wall shows a stranger nothing; a shared kit is the natural way a runner shows dialed.run to someone new. Time and place stay off because a daily post at the same hour narrows a runner to a city and a routine.
D-152 A world kit's link previews as its own card (owner, 2026-10-25; proposal 163): round 26 #22's entry card, revived from R-150, with D-144's Visual Crossing credit (R-160), and never the photo. Every other link keeps the default card. Amends D-107. D-107's default card for every link. Most people who see a link in a thread never click it; the card is what they see. D-107's reason (a preview must not carry what a signed-out visitor may not see) does not apply to a kit its runner made public.
D-153 Photos on world kits are published to a public R2 bucket on a custom domain and served from Cloudflare's cache, under an unguessable key, deleted and purged on any hide, takedown, ban or audience change (owner, 2026-10-25; proposal 163, R-153's design). Cloudflare's CSAM Scanning Tool covers them there, so world photos do not wait on PhotoDNA; D-129 stays the plan for upload-time matching of every photo. The bucket, the domain and the tool are owner-made. Amends D-69 for world photos only; runners photos stay private behind sign-in. D-69's "no photo is served with a public Cache-Control" for world photos. A world photo has no audience to protect, and serving it from the cache is what lets Cloudflare's tool scan it.
D-154 Public URLs never vary by viewer, and are cached at the edge for 60 seconds (owner, 2026-10-25; design 163). /kit/<entryId> (renamed from /k by the owner, 2026-10-29) is a public kit and /@handle the public profile; the signed-in profile (H) moves to /feed/u/<handle>. Every viewer of a public URL gets the same page, a signed-in one with a link into the app. The 60-second TTL bounds how long a retracted page can stay up; a purge by tag on retraction only shortens it and never fails the write. No static or eventually consistent build. /@handle as the signed-in profile (round 26 #7, FEED-10). A page that never varies by viewer can be cached with no risk of serving one runner's view to another. A static build would make every write a re-render trigger, and a missed trigger leaves a retracted kit public with no time limit.
D-155 A session that lapses mid-form carries the runner to log-in and back to the form's page, and does not restore what was typed (owner, 2026-10-31; design 166). Every signed-in form carries: a run, garment or settings form names itself ("Log in and you're back on your run."), and every other one says "Log in and you're back where you were." The second sentence is always "What you typed wasn't kept." Signed-out forms carry nothing. Lane: ui/use-form-submit (carry), ui/session-lapse, lib/carried-form, modules/auth/auth-page.tsx. The Form Contract's SESSION EXPIRED ("route to sign-in carrying the payload, return to the filled form") and Au7's "What you entered is kept." (design-deltas item 93). The same call as D-96 and D-102 for the terms refusal: restoring needs every form to hand its values over and take them back, a rule for a photo mid-upload, and typed notes kept in session storage. Saying so before it happens costs the runner nothing they were not already going to lose.
D-156 Weather attaches only finished hours: an hour is fetched once it has settled, two hours after it ends (owner, 2026-10-31). A run saved or imported sooner attaches the hours already settled, so its conditions show at once over the hours that are in; the rest stay owed, the run keeps weather_status = 'pending', and the hourly weather-retry cron fills them once they settle (law 8c reconciliation). A run wholly inside the margin waits for the cron, as it did before #288. The cron respects the same margin, and its five-hour give-up window opens when the run's last sampled hour settles (or on arrival, for a file uploaded later), so a long run is never given up on before its last hour can be asked for. A start-time correction to an hour not yet settled keeps the move and leaves the run pending, rather than reverting as if the weather had failed. The margin is SETTLE_MARGIN_SECONDS in modules/weather/attach.ts. #288's attach at save time, which asked for a run's final hour as soon as it was saved; and the cron, which could already cache an hour still in progress. Visual Crossing's value for the current or just-finished hour can still be partly forecast. The observation cache is shared — every runner in the rounded cell that hour reads the same row — and never refreshed, so whatever the first fetch got becomes that hour's weather for everyone, permanently, and the training signal the Call will learn from. Two hours is the margin the owner accepted for station observations to land. The cost is a wait of up to about three hours for a just-finished run's conditions, or its last hours' part of the span.
D-157 A photo still pending screening 24 hours after upload goes to the Desk's review queue for a person (owner, 2026-10-31; design 167). Garment and entry photos alike. The 24 hours run from the upload, read from the ULID it minted (an entry photo's id, a garment photo's version), so no column is added. The screening-retry sweep opens the row, with source unscreened (the classifier never answered) or missing_file (storage has no object), and the hourly Sentry line names each photo still pending by id, marking the missing ones. A row stays one per photo; one a person settled within the day is not re-opened, and a screening that finally passes settles it. A person's Approve on an unscreened row publishes the photo (pass), and a missing_file row offers Clear photo instead of Approve, which takes the photo off its kit or garment, cites no rule and asks its owner in the bell to add it again, a bell row only with no email; Remove and Remove as suspected CSAM stay on that row, because a missing file can still be a report on real abuse (owner, 2026-10-31). Leaving such a photo pending indefinitely, with an hourly count of photos and no ids as the only trace. Law 6: a missing object, or a photo the classifier will not answer for, never heals by retrying, and its owner saw "being checked" with no end. A day is long enough for an outage to clear and the retry to win, and short enough that a runner is not left waiting on nobody. The Desk already decides photos, so it needs no new surface.
D-158 Strava refresh tokens are stored without app-level encryption (owner, 2026-11-06). strava_connections.refresh_token and strava_revocations.refresh_token stay plain text in D1, which encrypts at rest. Nothing wraps them in a key the Worker holds. Encrypting the tokens with an app-held key before they are written. A refresh token is useless without STRAVA_CLIENT_SECRET, which lives only in the Worker's secrets, and since D-54 the token does one thing, which is to revoke. A key held by the same Worker that reads the column adds a secret to rotate and protects against nothing D1's encryption at rest does not already cover.
D-159 Task 090's validation study is cut (owner, 2026-11-06). No hand-validation study runs, before or alongside v1. The Call epic starts from the generic dressing chart and learns from v1's verdict history. Supersedes D-02. D-02: the study runs in parallel and its findings tune the Call epic. The Call epic's engine already starts from a published heuristic floor and earns its confidence from runners' own verdicts (docs/post-mvp.md), so a separate study would hold up the epic for a signal v1 collects anyway.
D-160 The per-item note is cut for v1 (owner, 2026-11-06; R-8). An outfit entry's item keeps its optional flag; the note beside it is no longer shown or read back, and no input writes it. The outfit_entry_items.note column stays (expand→contract, law 8) until a later contract step drops it. R-8's "wire the input or drop the field". No screen writes a note today, so showing one read back is a path with no input. Flags carry the per-item signal; a note can come back later with an input of its own.
D-161 Intimate-image takedowns: each case is judged, and "intimate" is the statutory line (owner, 2026-11-06; R-147). (1) An upheld intimate-image takedown flags the poster's account on the Desk for an operator to review. No account is closed or banned automatically; the terms' serious-breach clause stays the operator's call. (2) The intimate reason, its Review row and a Decline's wording use the definition in 15 U.S.C. § 6851(a)(5) (nudity or sexual content). A photo of a person who did not agree to it, and is not intimate under that line, goes through an ordinary report, and an operator may remove it as a courtesy. R-147's open questions: an automatic consequence for the poster, and a wider "intimate" than the Act's. No law requires action on the uploader's account (the TAKE IT DOWN Act asks only for removal and reasonable efforts on known copies), and one upheld report can be a mistake or a dispute, so a person decides. The statutory line keeps the first-report hide (D-117) for the images the Act covers, so an ordinary post-run photo is not hidden on one report.
D-162 Automatic holds tell the author too (owner, 2026-11-06; design 152). When the classifier or a removed-image fingerprint match (D-147) holds a kit or photo for review, its author gets the same in-app notice a report's hide sends (hiddenNotice), with the reason "automatic screening" and the same appeal path. Suspected CSAM stays silent. Amends D-145, which covers a report's hide only. D-145's notice on a report's hide only. DSA Art. 17 covers any restriction of visibility, automated ones included, so a hold by screening owes the author a statement of reasons the same as a hold by reports does.
D-163 A runner can correct a shared product's facts, and the Desk applies or declines it (owner, 2026-11-06; R-74, rounds 48 #129–#130). A product_corrections table holds one correction per runner, product and field; §AG's WRONG? link opens the sheet, and the Desk's Corrections queue applies or declines it. Before friends. R-74's 2026-09-19 call to wait for task 110. Owner call. The Desk shipped without a correction queue, so the wait for task 110 had no end.
D-164 Opening a review's who-reported fold is recorded in a review_audit table (owner, 2026-11-06; R-66). Before friends. Logging nothing until an audit table existed. Owner call. Round 8's D1 says opening the fold is recorded; Sentry is for errors, not an audit trail.
D-165 The attach-kit prefill matches on a run's worst hour, the same hour its conditions use (owner, 2026-11-06; design 054). Matching on the starting hour (design 054 left prefill there). Owner call. Run conditions already use the worst hour (design 054).
D-166 A product link typed with an explicit http:// is accepted and upgraded to https:// (owner, 2026-11-06; design 042). Refusing http:// (the default design 042 left). Owner call.
D-167 The kit caption is cut for v1 (owner, 2026-11-06; PR #277). No input writes it; the column stays (expand→contract, law 8). Building a caption field for #277's display. Owner call. No form ever wrote a caption.
D-168 Follow notifications ship before friends; "useful" never notifies (owner, 2026-11-06; design 129, round 48 #131). S1's '@x found your run useful' row. Owner call.
D-169 "Worn by N runners" waits for the public stage, and shows only above a minimum N (owner, 2026-11-06; task 101). Counts still derive only from shared entries (D-29). Lane 104 computing it for v1 (task 101). Owner call.
D-170 A removed public photo may be served from cache for up to 60 seconds (owner, 2026-11-06; PR #277). No instant purge by API token. An instant purge (a new secret). Owner call. The privacy policy says "up to a minute", and D-154's TTL bounds it.
D-171 No staging environment; production deploys get guard rails instead (owner, 2026-11-07; #419). Gradual versions, a production smoke test against the uploaded version, and a rollback runbook before friends (#504); automatic rollback on error rate at public (#505). A staging Worker with its own databases and Strava app. Owner call. At this size the risks that matter (a privacy leak, bad data) are caught by tests before deploy or need a data restore, which staging doesn't give. A second stack costs more to set up than the guard rails.
D-172 R2 photos are backed up daily to a second bucket, from the first deploy (owner, 2026-11-07; #420). The daily cron copies new and changed objects. The owner creates the binding; a lane writes the code, its law-10 conformance test and the deploy-checklist row. The deletion tombstone as the only recovery. Owner call. Friends' photos are covered from day one.
D-173 Public launch stays invite-only (owner, 2026-11-07; #371). IS_INVITE_ONLY stays true; public is a wider beta with codes the owner hands out, and requesting access (#367) is the public way in. Open sign-up at public. Owner call.
D-174 Each database's migrations are squashed into one baseline before the first deploy (owner, 2026-11-07; #506). Columns already cut go first, so the baseline never has them (outfit_entry_items.note and the export's note column, #408). Deploying the full migration history. Owner call. Nothing is deployed yet (D-139), so the history carries no data.
D-175 Each run keeps a copy of its judged conditions in core, built before public and after the squash (owner, 2026-11-07; #409). The consensus reads the copy instead of walking to the weather database per run. Keeping the per-run cross-database walk. Owner call. The walk grows with every public entry in the window; the copy is one indexed read. The weather database stays split for its own reasons (docs/architecture.md).
D-176 No shop images: a garment or product shows only the runner's own photo (owner, 2026-11-07; #472). Enrichment stops copying shop images, products.image_key stops being written, and the stored copies are deleted before friends. Amended (owner, 2026-11-09): none were ever stored, because the app had never been deployed, so the purge script was removed as dead code (#542), and #555 dropped products.image_key. Copying shop product images. Owner call. Shop images return only under a licence that grants image use.
D-177 strava_revocations stays apart from the generic outbox (owner, 2026-11-07; #435; design 127's default). Folding it into the outbox. Owner call. Its rows hold a credential, and the generic outbox forbids a secret in a payload.
D-178 The name dialed.run and its brand are reserved by a trademark policy (owner, 2026-11-08; TRADEMARKS.md). Forks may use the code under the AGPL, but may not run it for other people as dialed.run or with its logo, wordmark or app icons. The code's licence is unchanged. Nothing: the name had no policy. The AGPL licenses copyright, not the name, and its §7(e) lets a licensor decline trademark rights. A clone passing itself off as dialed.run is the realistic harm a public codebase invites, and only a trademark policy answers it.
D-179 The production smoke account is invisible to runners (owner, 2026-11-09/10; #533, design 193). It's marked by user_profiles.system_account, which no app code writes; the owner sets it with wrangler d1 execute at deploy. Profiles, search, feeds and every count exclude it, enforced by one gate per surface and an architecture test. Handle smoke_test, invite issued from the Desk, password rotated yearly or on a suspected leak.
D-180 Backups outlive a delete by at most 30 days (owner, 2026-11-10; #529). Photo backup copies are deleted 30 days after the original, matching D1 Time Travel's 30-day history. The privacy policy says so.
D-181 Local Stryker concurrency stays at 4 (owner, 2026-11-10; #558). 6 measured about 13% faster on a gate, but it isn't worth the machine load during a wave.
D-182 Workflows pin actions by version tag plus a lockfile (owner, 2026-11-11; #601). Every uses: names a version tag, and .github/workflows/actions.lock records the commit each tag resolved to. SHA pins in the workflow files go, since GitHub's lockfile makes them unnecessary and Dependabot keeps lockfile entries current only for workflows that use one.
D-183 The guardrails init --plan drift is declined (owner, 2026-11-11; #556). Nothing from init is applied. The divergences this repo expects are listed in docs/guardrails/init-drift.md; a divergence not on that list is a finding.
D-184 Nothing is filed with third-party projects (owner, 2026-11-11). Agents don't open issues, PRs or discussions on other projects' trackers, our tools' upstreams included. A problem found in a dependency goes to the owner as an issue here.
D-185 A control's kind follows when it saves (owner, 2026-11-13; Round 54 · 166.3). A setting that saves the moment it changes is a switch, announced "Saved". An option that waits for a submit (a form field or a consent) is a checkbox. One of several, waiting for Save or Next, is a radio. Filters stay chips. No control uses the browser's native look.
D-186 Each surface has one primary button kind (owner, 2026-11-13; Round 54 · 167.1). Product flows use the pink caps button; account, signed-out, settings Save and confirm sheets use the ink pill; the shell and marketing use the pink pill. Secondary is always the hairline pill, tertiary a text link, and a hairline pill is never alone. That rule is per surface: a surface whose only button is secondary in kind, like P2.5's Save, keeps the hairline pill (owner, 2026-11-13; Round 56 · W1/229).
D-187 A control's accessible name is its visible words (owner, 2026-11-13; Round 56 · W1). Anything more, such as a consequence or a handle, is a shown description tied to the control. Icon-only controls and screen-reader-only descriptions wait on #662 (Round 57).
D-188 A profile report has its own reasons (owner, 2026-11-13; Round 55 · 214, #636). Reporting a runner's profile offers its own list: their name or photo is offensive, they're pretending to be someone else, harassment, spam, other. The hide sweep counts only reports on what a report can hide, entries and photos, never profile reports.
D-189 A row that scrolls is a region a keyboard reaches (owner, 2026-11-14; #659). A swipe row with nothing focusable inside is a labelled region with tabIndex 0, so keyboard and screen-reader users can scroll it, rather than a gesture with no scroll region. ESLint's no-noninteractive-tabindex allows region for this.
D-190 A band the runner set reads the same wherever it shows: SET, the band, and its sky (owner, 2026-11-14; #704 R2-RUNS-04). The shared Conditions contract carries an optional sky for a set band (source: 'manual'), and the weather reader passes it on rather than dropping it. So the backlog reads "SET · 47–56° · DRY" as the runs list does, never the stored middle, "MANUAL" or a 0 mph wind.
D-191 Copy that counts agrees with its count (owner, 2026-11-14; #704 R2-RUNS-07, #659). With one, the sentence is singular ("Your run, its outfit and verdict"); with two or more it names the number and goes plural ("All N runs, their outfits and verdicts").
D-192 The fingerprint breaker counts distinct uploaders (owner, 2026-11-16; SA-M20, design 232). The hour's count is of distinct uploaders, not rows: a kit photo's author, a garment's owner, or the resolver once the subject is gone. Amends design 164's fallback.
D-193 A speed check sends at most 5 findings to Sentry a firing (owner, 2026-11-16; SA-M34, design 233). The first 5 go as their own events, and the rest as one event that counts them. Amends design 165.
D-194 The app moves to app.dialed.run before the first deploy (owner, 2026-11-16; SA-L08, design 233). D-99's move is done now, not in the deployment sweep: the Worker answers on app.dialed.run alone as a Custom Domain, workers_dev and preview URLs are off, and BETTER_AUTH_URL is https://app.dialed.run. The dashboard steps are deployment-plan rows. Amends D-99.
D-195 A closed address can't make an account, and the email form doesn't say so (owner, 2026-11-16; design 228; Round 60 · 328). Email sign-up answers a closed address as it answers a registered one. Google sign-in returns to Au2 with ?error=ADDRESS_CLOSED and the band "This address can't be used to make an account."
D-196 A closed account waiting on deletion reads as both on the Desk (owner, 2026-11-15; design 228; Round 60 · 327). In its deletion week it shows [CLOSED] [PENDING DELETE], and the digest says "N account deletion(s) wait for a copyright hold to settle".
D-197 A manual run's title is the runner's own (owner, 2026-11-16; Round 60 · 330). Other runners see the day and distance on W1, feed cards and kit detail. R1's title field says "Only you see the title."
D-198 Blocking has no ceiling and doesn't stop a report (owner, 2026-11-16; Round 60 notice). Settings lists every blocked runner, unpaged, and a runner the poster blocked can still report the kit.
D-199 /api/health tells a stranger only whether the app is up (owner, 2026-11-17; #824, design 237). Anyone gets { ok, version, at } with 200 or 503. A signed-in operator (the Desk's gate) also gets checks, missing, staleCrons and a set-or-missing line for every deploy value.
D-200 The daily digest names only the deploy values the plan requires (owner, 2026-11-17; #824, design 237). It lists a missing value only when the deploy plan marks it yes, strongly, before friends or before public. Optional, owner-only and for-Strava values show only in the operator's health report.
D-201 Modules split into sub-modules that meet at their doors (owner, 2026-11-18; #796, design 240). Another module may import a sub-module's index.ts. dispatch (the cron and queue dispatch) is the only new top-level module; ops' floor (the outbox, Sentry, Turnstile) stays in ops as sub-modules. Components move with their service code. The purge stays in account as account/purge. A ratchet holds each module to fan-out 7, door width 45, nothing past the door and 50 files, with pins that only fall.
D-202 An import is claimed with a token its sender mints (owner, 2026-11-18; #841, design 242). imports.claimed_by holds it, and a redelivery takes back a processing row only where the token matches; a message without one takes back only a row with none.
D-203 The friends freeze, and minor Design asks during it (owner, 2026-11-17/18; #766, #850). The remaining friends build merges, then only fixes merge, then the third audit round is pinned at the freeze commit. A new minor Design ask still goes to a round but takes the soon-after-launch milestone unless it blocks a friends item; the lane builds to the contract or the earlier ruling meanwhile, and a ruling that changes the build is a new build issue.
D-204 Security fixes ship quietly and deploy at once (owner, 2026-11-18; #788). A security fix's design doc lives in the private home, its PR body describes only the change with no audit id tied to an area, and each security merge deploys straight away. A private branch and deploy path is added only if the third audit round finds a high.
D-205 /open-source names each licence group by its published full name (owner, 2026-11-18; #831): "Projects under the MIT License", "Projects under the Apache License 2.0", and so on.

Open items deliberately not decided here: the call's model design (see docs/post-mvp.md), full-E2 ranking, city-launch selection, comment revisit criteria.