Availability
All career features ship together and are available without a release flag: Today, the coach, Knowledge, offer comparison, the application workspace (fit checks, application answers, interview preparation, practice, debriefs and employer messages), discovery, reminders, and recorded voice practice. AI generation requires a usable AI connection. Voice can use a speech-capable connection or supported browser speech. Scheduled work requires a working deployment scheduler and explicit user opt-in. The verification section below covers deterministic checks and live-provider evaluation before production deployment. The assistant uses the existing saved AI connections. ConfigureENCRYPTION_SECRET, then add and test a provider in Settings → AI & developer. Users choose their provider; provider usage is billed under that connection. Public research uses the selected web-access connection. See web access and job search and AI configuration for deployment and credential requirements.
Posting capture
Browser capture and native share/import feed the existing application review and save service. Updateadding-an-application, editing-on-mobile and self-hosting/job-search-and-ai with capture changes. Fictional screenshots belong under docs/images/guides/career/ (posting-capture.png and native share review when available). Validate the actual receive target on each platform; a web preview or paste form alone is not native share evidence.
Employer-form packet prototype
applications/packet.ts extends existing private-profile, selected-document/export/version and reviewed-answer owners. Review projects only opted-in contact fields, visible work history, letter text and explicitly selected eligible answers. Approval is bound to the exact URL and keyed source fingerprint. Existing owner locking revalidates application, linked letter source, profile, documents and evidence before and after rendering; changed or unavailable sources fail without saving packet versions. The existing version owners retain named snapshots. Packet creation never records submission or changes stage, and makes no AI call.
Both authenticated routes require read and write permissions. The private packet routes are absent from the MCP allowlist. No server fetch of the employer URL occurs. Combined PDFs are limited to 2 MiB. The JSON packet contains personal data and exact PDF bytes; downloaded copies cannot be revoked. Account export includes document histories (document-versions.json in the web ZIP, documentVersions on native); existing account/document cascades remove stored versions. No new table or migration is needed.
The desktop posting-capture helper uses existing activeTab scripting permission to fill an imported, explicitly reviewed packet. Packet data stays in popup memory, never extension storage. Capture behavior and permissions remain unchanged. Only the exact reviewed URL can run packages/utils/src/form-fill.ts. Native mobile retains packet review, copyable answers, approved PDF downloads and opening the form in the system browser for manual completion. Android/iOS autofill is excluded by the owner’s explicit decision; no embedded browser or WebView dependency ships.
Every desktop fill preserves existing values, emits no input/change/click/submit events, checks file hashes and names exact version IDs. Framework controls, custom controls, ambiguity, field limits, account/captcha and additional pages have manual handoffs. Reports distinguish filled, skipped, manual and error; DOM file presence does not establish ATS upload acceptance. Do not extend this support matrix without controlled proof. Current real ATS compatibility remains unverified.
Use disposable PostgreSQL and synthetic users for account-backed proof. Follow the browser-test environment recipe; run:
RUN_CAPTURE_EXTENSION=1 for the actual Chromium helper case; otherwise that case is intentionally omitted. The shared-engine spec launches Chromium and host WebKit to check the desktop script. The earlier loopback lab under tooling/autofill/ remains historical experiment evidence and does not verify the new product integration.
Mobile packet changes reach users with the next store build; the app has no OTA updates. Deploy the additive backend first; old mobile APIs stay compatible. The owner-approved native-autofill exception is listed under “Parity exceptions” in apps/mobile/AGENTS.md.
Refresh writing-application-answers, career-workspace-reference and exporting-your-data with packet changes. employer-form-packet-v2.png is an actual fictional-account product screenshot; it makes no native-device claim. Keep the original employer-form-prototype.png as historical lab evidence.
Where things live
The career platform has two places:- Career (sidebar): Today (notifications that need you, new roles from discovery, the next interview and your schedules), Coach (conversations across your whole search), Knowledge (facts, STARR stories and preferences) and Offers (two saved offers side by side).
- The application workspace (open an application, then Open workspace): one job in seven tabs, in the order a search happens. Fit, Apply, Interview prep, Practice, Debrief, Messages and Library. The workspace opens on the tab for the application’s stage, and Ask the coach opens the coach beside any tab. A job can have several conversations: start a new one, or reopen or delete an earlier one from the sheet’s history.
Prepare for an interview
- Add the round with Edit in the workspace’s Next step card: type, date and time (in a chosen timezone), length, who you’ll meet, and what they asked you to prepare. A linked resume is optional.
- On Interview prep, choose Plan my preparation. The briefing is a timed plan, questions to ask, the stories to bring and a 10-minute briefing (your focus, an opening line, two stories and what isn’t evidenced yet).
- Switch on Refresh the briefing 24 hours before to run it again automatically. Editing the interview later moves that run with it; when the new time is already inside the lead, the refresh is paused instead of running at once.
- From 15 minutes before the start, Interview prep becomes a short checklist that needs no AI or research, with Rehearse one answer aloud.
Control knowledge and what is sent
Knowledge holds facts, stories and preferences separately from document formatting. Add a few relevant accomplishments, or tell the coach about them; no questionnaire is required.- Remember facts from my messages, at the top of Knowledge → Facts on web and native, has three modes. Ask first (the default) shows a proposal under the coach’s answer; nothing is saved until you choose Remember. Automatically saves facts you state and shows Undo. Never turns memory off. Mock answers are never remembered.
- Facts show where they came from and their correction history. Correct keeps the original in the history and marks stories that use the fact for review. Switching a fact off stops the coach from reading it; a switched-off fact can’t back a story. Facts private to one application are used only when coaching for it, until you choose Use for all applications.
- Forget removes the fact after a six-second undo, unlinks it from stories and saved items, and keeps a content-free marker so it isn’t re-extracted from the same source; a later reviewed resume import can still re-add it. The original resume, message or file isn’t changed.
- STARR stories contain Situation, Task, Action, Result and Reflection with the facts that back them. Situation, Task and Action are required; leave Result and Reflection blank until you can support them.
- Preferences (target roles, places, what matters, minimum base salary, office days, notice period) feed discovery, the match chips on new roles, Fit’s requirement checks and Offers. All times in Career are shown in the account’s timezone, set in Settings → Preferences (the device timezone for new accounts; existing accounts that never changed it stay on UTC).
Use the workspace tabs
Add offer in Offers or an application’s Messages tab saves structured terms manually without an AI connection. Offers compares two saved offers in their own currency and pay period. Differences are marked only when the currency and period match, unknown or conditional values are flagged, and totals aren’t added up. Plan what to ask opens the coach with both offers as context.
Undo replaces confirmation dialogs throughout: forgetting, deleting, dismissing and applying changes all show a six-second Undo.
Operate scheduled work
Supported deployments schedule career work automatically once a minute. Users still explicitly enable each schedule and configure any required AI/web connection. Preparation stays pinned to its selected provider; it does not fall back silently. Editing that provider’s key, model or address pauses its schedules until renewed approval.
Vercel’s checked-in once-per-minute schedule requires Pro or Enterprise. Hobby supports at most one invocation per day and rejects more frequent schedules during deployment. On Hobby, change the schedule to
0 0 * * *; daily ticks can delay reminders and preparation by a day. See Vercel Cron limits. Native Vercel Cron runs on production deployments. Cloudflare trigger changes can take up to 15 minutes to propagate; see Cron Triggers.
Generate the Vercel token with openssl rand -hex 32, supply it through the deployment’s secret mechanism, and never log authorization headers. The UI checks deployment configuration, not successful delivery of ticks. Verify a due test reminder in a dedicated installation after deployment. Minute-by-minute polling queries PostgreSQL even when no schedules are due, so a database that scales to zero can remain awake.
All adapters await the same bounded PostgreSQL-backed processor. Node prevents overlapping ticks within its process; database claims coordinate multiple processes and native platform invocations. Cloudflare keeps scheduled work alive with waitUntil() and releases its invocation’s database pool after work completes. No task depends on an open browser.
Optional manual invocation
An operator can trigger a tick for diagnostics or use a custom scheduler. SetCRON_SECRET (at least 16 characters), then call GET or POST /api/career/run with Authorization: Bearer <secret>. Without this token the HTTP endpoint returns 404; invalid authorization returns 401. Internal Node and Cloudflare ticks continue without this token.
From a source checkout with Node.js 26, a manual tick can use a protected environment file outside the repository:
Enable a user’s schedule
In Career → Today, Schedules → Add (or a schedule’s row) opens the schedule sheet:- Find roles: a public search (it starts from Preferences) run every day, every 3 days or every week. Results wait in Today for review; tracking one creates an application at Saved only when the user chooses Track in Applications.
- Interview prep: an interview, when the briefing runs (24 hours before, 2 hours before, or Morning of: 08:00 on the day, moving earlier to two hours before interviews starting before 10:00) and a tested AI connection. If the interview starts sooner than that, saving runs it now.
- Remind me: the application’s single follow-up date and note; it fires once into Today. Application edits and Today edits share the application-owned state, with a schedule as delivery metadata. Delete clears both after Undo; pause retains the date but stops delivery.
query contract. Remount the draft owner when accounts change.
Application follow-up edits preserve the reminder’s existing wall-clock time, including note-only edits. Exact-time reminders retain their instant when the account timezone changes. Legacy date-only follow-ups retain their selected day at 09:00 in the current account timezone; changing that timezone updates pending delivery in the same database transaction. Queued or running work for the previous clock is canceled and replaced at the new instant. Completed occurrence receipts follow the logical day without re-delivery; completed jobs retain their original actual due times and notification/email identities. Note-only edits update pending publication under the existing locks and never rearm a completed occurrence.
The migration marks legacy UTC-midnight date-only values without rewriting their raw timestamps; explicit Today times, including midnight UTC, stay exact. Duplicate legacy reminder rows and their job history remain in account exports; only the newest delivery row is shown and enabled. The account clock trigger reconciles pre-existing logical-day delivery indices and runs for both auth update paths. A timezone containing no such local day is rejected atomically.
Dates and times are entered in the account’s timezone. Recurring discovery preserves its local time across daylight-saving changes; a local time that doesn’t exist during a clock change moves forward to a valid time. Missed intervals are coalesced rather than replaying every missed scan.
Pause or delete a schedule to cancel its queued or running work. Rescheduling an interview moves its briefing schedules to the new time and cancels work queued for the old one. A schedule whose new run time has already passed is paused, never run immediately. Removing the interview pauses them and adds a notification to Today. Cancellation prevents later result publication but can’t undo provider calls already made.
Interview calendar files use the shared UTC ICS builder in packages/utils and interview mapping in packages/schema. Web and native save notices offer Update your calendar; committed deletions offer Download cancellation. Both keep the round’s original UID. Exports include METHOD and epoch-second SEQUENCE; cancellations add STATUS:CANCELLED, advance a same-second sequence and omit VALARM. Undo must never export cancellation. Interview deletion preparation loads and registers the canonical application detail reader before acknowledging the action, including Calendar summaries without a detail query or hydrated without a query function. Undo and Retry must read that owner before revealing retained data; they never write an old interview back. These files require manual import into an external calendar; calendar-client import behavior is a separate compatibility check.
Understand delivery and retry limits
PostgreSQL owns due work, leases, attempts, and deduplication. A tick handles at most two jobs, and active work is serialized per user. Duplicate ticks and reclaimed leases are designed to avoid duplicate visible results. Failed jobs wait five minutes before a later tick can retry, with at most three attempts. Retries can repeat provider work and incur usage. Scheduled preparation requires its explicitly selected AI connection and never falls back to a default. Deleting or switching off that connection, or changing its credentials, model or address, pauses its schedules and cancels queued/running job leases under the same owner lock. Final authorization checks the selected connection and active lease before transmission; an already-sent request cannot be recalled. A legacy preparation schedule with no selected connection is paused with an actionable in-app failure. The user must choose and approve a tested connection before enabling the schedule again. A generation failure follows the normal retry schedule. Exhausted work produces an actionable in-app failure. The recipient revision approved at enablement is stored internally, including for server-managed connections. The scheduler rejects a different key, model or URL even when the connection ID stays the same. Migration20261031234857_ancient_rhodey pauses existing preparation schedules and cancels their queued/running leases because those rows have no approved revision; discovery and follow-up work are unchanged. Users review and enable preparation again through the existing approval flow. Changes to a managed connection refuse transmission and follow the bounded failure/retry flow until the schedule is reapproved. A rollback of the new column must not automatically re-enable paused schedules.
Discovery reads each new search result and keeps it only when it is one open role’s posting: the page describes exactly one schema.org JobPosting whose closing date hasn’t passed, or it sits at a posting address on a common applicant tracking system or job board (Greenhouse, Lever, Ashby, Workable, SmartRecruiters, Workday, LinkedIn, Recruitee, Personio, Teamtailor, Wellfound or Welcome to the Jungle). Job-board lists, careers pages and articles never reach New roles and aren’t read again. Pages are read with the built-in reader first and with the web search connection only when that fails. URLs are canonicalized and common tracking parameters stripped, so a result seen before stays quiet. Untracked visible postings found again are rechecked after a week to refresh their description and open status without another new-role notice. Tracking a role saves its posting text, company and location with the application, so Fit can check it at once. Dismissed results stay hidden. A posting being found is an observation, not proof that it is open or genuine.
Owner-email notifications require explicit opt-in, a verified account email, and working SMTP configuration. In-app results are retained independently. The email attempt is recorded before contacting SMTP, with a deterministic message ID. A claim left unfinished by a stopped worker becomes eligible again after ten minutes. Completed sends and recorded failures are not retried. A crash after SMTP accepts a message but before completion is recorded can cause a duplicate; SMTP acceptance does not establish inbox delivery. A recorded failure appears in-app; use the in-app result as the durable record.
Notifications are written in English for now. Each one is stored as what happened (its type and details), and its words are written when it’s shown, so a later release can show notifications in your language and their times in your timezone.
Optional voice practice
On Practice, choose Speak. Voice starts with the default connection when it supports speech, otherwise another usable speech connection; Voice settings chooses another, or a different model.
Speech calls are billed separately from text coaching and may use tokens on the connection’s account; the page says so where they happen. Without a connection for a job, the browser does it: its own voice reads questions aloud (every major browser), and its dictation turns your answer into text (Chrome, Edge and Safari; not Firefox). Browser speech is free, but depending on the browser, its maker may process the audio. Where neither is available, typing remains a complete alternative.
With a connection, record up to two minutes, or choose a nonempty WebM, MP4/M4A, MP3, WAV, or Ogg file of at most 2 MiB. With the browser’s dictation, your words appear as you speak, for up to two minutes. Check the transcript (it becomes your answer and can be edited), then choose Get feedback. Original and edited transcripts remain distinguishable and can be deleted.
The server keeps the transcript, not the raw recording. After successful transcription, browser audio is cleared unless you choose to keep it locally for download. Failed transcription leaves the current recording available for retry. Local audio is temporary and clears when the panel unmounts. Provider retention follows the selected provider’s policies. Spoken questions send only the question’s text.
Migrations, export, and deletion
Career storage ships through additive migrations, including new user-owned tables and conversation context fields. Back up your database and file storage, review generated SQL, and apply the normal migration workflow before deploying the update. Server startup also applies migrations. Account export includescareer.json: preferences, facts and source/revision metadata, stories, reviewed answer library entries (scope, source, evidence revision references and review time), saved items, workspace state (answers, checklist, plan ticks), schedules, notifications, discovery results, jobs, voice transcripts, and career conversations/messages with attachment metadata. Conversations include their application context. Provider credentials are not included. Attachment metadata is not a backup of the attachment’s binary contents.
Account deletion removes the user’s dependent career records along with the existing account data and private upload storage. Deleting an individual fact, saved item, transcript, application, or conversation has narrower effects; it does not delete copies previously exported or sent to a provider. Include database backups and externally retained data in your installation’s retention policy.
AI request usage
The authenticatedai.usage procedure (GET /api/openapi/ai/usage) reports the account’s current Monday-to-Monday half-open week, calculated in its saved timezone with PostgreSQL calendar arithmetic. Both clients expose it in AI settings without requiring a live provider or per-operation sharing approval: reading server records performs no AI action.
Inference factories require server-created account/action/operation scope. The transport reserves one durable ai_request row before each provider transport invocation, after account, context, provider availability and abort checks. SDK retries and later model calls in tool loops reserve separate rows; stream chunks, tool execution, replay and local saves do not. Failed or cancelled invocations remain counted. Network policy remains owned by the existing transport. A crash between reservation and I/O may leave an unsent attempt, so totals do not assert remote receipt, billing, tokens or cost.
Connection labels, provider types and requested models are captured when the operation is created; later connection edits or deletion do not rewrite them. Saved artifacts, reusable answers, transcripts, writing reviews and conversation metadata reference server-created operations. Client-referenced answer operations must exist for the same account and answer action; legacy answer writers retain recorded sources for stable answer identities. Counts and model summaries are derived from owned request rows rather than client claims. Existing results without references stay unknown.
Account export retains usage and writing reviews within the existing Career backup object, which older archive builders preserve. Account deletion cascades usage with its user. Never point metering integration fixtures at a production database. Verify actual SDK transport retries/tool loops/streaming, account isolation, recipient changes, week/DST boundaries, saved result reload, connection removal and account export/deletion independently of UI presentation.
Verify before rollout
Use a disposable PostgreSQL database and synthetic users. These integration suites create an isolated schema, apply the complete migration chain, and remove their schema afterward. Never point test environment variables at production data.pnpm test:db exec creates and migrates disposable databases for your checkout on the local PostgreSQL from compose.dev.yml, then runs the command with the test database variables set. It also sets REQUIRE_DB_TESTS=1, so the career database suites fail instead of skipping when a database variable is missing; a green run with skipped suites is not verification. Follow development setup for details and the remaining package prerequisites.
The deterministic cases cover ownership, submitted-version isolation, correction/forget behavior, practice/debrief separation, stale material, exports/cascades, overlapping ticks, retries/cancellation, recurrence, and endpoint authorization. Also run affected package typechecks, existing assistant regression suites, boundary/lint/format checks, and the production build. Test the deployed endpoint and owner-email flow in a dedicated installation as described in deployment checks.
Career behavior is covered by the database and unit suites above; the browser suite keeps only cross-system journeys (see the browser test README).
Opt-in live evaluation
Live evaluation is a manual, separately authorized use of configured provider accounts; the deterministic suite makes no live quality claim. Use synthetic histories, known public sources, and a recorded provider/model configuration. Include several professions (for example software engineering, nursing, and sales), junior and experienced candidates, and at least three supported UI/coaching languages. Include career changes, missing metrics, conflicting documents, and ambiguous interviewer identities. For each configuration, record the following outcomes without retaining private resumes or transcripts in operational logs:
Treat any cross-account access, invented personal claim, wrong submitted document, or inability to remove private context as a release blocker. Record unresolved quality and accessibility findings before production deployment. Monitor latency, completion/failure counts, provider usage, corrections, and user-rated usefulness; avoid raw document, research-page, or transcript content in operational logs.
Private application profile boundaries
Application-only contact, address, links, country-specific authorization/sponsorship, relocation, requested salary and references live inapplication_profile, outside career_profile.data. Career AI context currently spreads Career preferences, so never put application-only fields there. Older whole-object preference saves cannot erase this separate record. Profile routes use authenticated identity, delegated read/write/delete permissions and owner-first revision checks. Account export includes the record; MCP account-export results explicitly exclude it. New profile procedures are not MCP tools.
Guest linking moves a profile only into an empty destination and changes its revision. Two profiles cause a conflict before ownership changes; the user must export and delete one. Account deletion cascades the row. Profile deletion does not alter shared preferences, documents, application answers or submitted snapshots. Self-identification storage is deliberately excluded for this phase; categories, retention, consent and prompt exclusion require independent review before any extension.
Mobile changes use JavaScript, copy and existing dependencies; no native dependency, permission, entitlement, native config or Forme pin changed. They ship with the next store build. Deploy the additive database migration/backend before that build.
Keep product guides current
Application removal uses durable Trash on web and native.application.trashedAt retains timeline, document/version links, workspace, scoped Knowledge, coach records and files. Active workflow queries use the application lifetime guard; full account exports intentionally include retained rows. Restore/permanent-delete controls carry the listed timestamp and serialize with owner writes. Application Trash has no automatic expiry; document Trash retains its existing 30-day policy. Update managing-an-application and using-the-trash, including fictional recovery captures, when changing this lifecycle.
Update user-facing guides in the same change as the behavior they describe. Start with the changed screen or service below, then follow its guide’s related links. The overview explains the product, the first application tutorial teaches the starting flow, and the reference records controls and limits.
Guide names above live in
docs/guides/; screenshots live in docs/images/guides/career/. Keep new pages in docs/docs.json. Use fictional people, companies, messages and results for screenshots. Capture the actual interface, inspect the image and explain what the reader should notice in its caption and alt text.
Verify labels in the web component and behavior in the owning API service before documenting them. Recheck opt-ins, default AI selection, chat versus workspace context, account timezone, the early-interview exception, voice fallback, unknown values and Undo. A fact is a user-confirmed record; it is not independently verified. Result fields can remain blank but do not enforce proof. Never promise fixed daily role counts, guaranteed inbox delivery or error-free AI claims.
Check local links, image paths, navigation entries and formatting after edits. Preview changed pages when a documentation preview is available. Changes confined to prose do not need application tests; changed behavior still needs its normal checks.
Discovery setup and explainable triage
Target-role suggestions extend the existing career API. The provider and context are explicit; generation returns transient drafts with actual source facts/quotes and stores no confirmed evidence. It rechecks shared Knowledge and resume context after generation so in-flight revocation fails. Neither application-only profile nor reusable answers enter the request. Manual target roles, search preferences and discovery require no AI. Review and save preferences before creating a schedule; existing queries are stable snapshots of setup text. Discovery retains the existing URL identity, validation, weekly recheck, suppression and durable runner. Nullable retrieval provenance distinguishes a successful posting check from later search sightings. Text chips are preference phrases and keyword overlap, not AI fit. Conservative English salary/workplace indicators favor unknown values; snippets do not establish these details. Track/dismiss share the existing owner mutex; tracking a dismissed untracked result conflicts, while repeated tracking returns the same application. Bulk triage composes the existing APIs and reports partial outcomes. Usediscovery.spec.ts against an explicitly disposable DATABASE_URL for source/unknown/bulk triage. Role suggestion browser tests use the local scripted AI provider with FLAG_ALLOW_UNSAFE_AI_BASE_URL=true. Controlled queue tests include duplicates, closed/list pages, transient reader failures and unchanged retrieval budgets. Synthetic measurements are regression evidence, not live freshness or hiring-quality claims. Role suggestions add one optional provider operation; generation’s actual SDK retries and provider pricing still apply.
The mobile change is JavaScript, copy and catalogs with existing native dependencies; no native config, permission, entitlement or Forme pin changes. It ships with the next store build. Native-device workflow verification remains a release gate; do not describe an Expo export as device proof.
Target-role recipient versions
aiProviders.list adds an opaque recipientVersion. career.suggestTargetRoles requires that version alongside the explicitly selected provider ID. Web and native send the version displayed for approval; native also checks it before and after its per-operation dialog. API and generated MCP contracts share the same required schema. A mismatch returns CONFLICT before generation. The request uses the captured approved runnable configuration, including credentials, without resolving it again.
The provider owner computes the version from owner, ID, provider type, model, full base URL and credential revision. Personal credentials use their existing encrypted revision; managed server AI includes its effective environment key through a keyed hash. Retests, label changes, default selection and usage timestamps do not invalidate it. No new table or migration is needed. Provider responses add this field; older mobile consumers and other existing AI requests retain their contracts.
RX98 integration seam: preserve this provider-owned version in setup/list/test responses and caches, including server-ai:<user>. Never synthesize it from labels, test timestamps or provider ID alone. New role-suggestion clients require a backend exposing it before rollout. Other AI operations are outside this versioned endpoint’s scope. RX-76 adds discovery_source after the main-95 migration. Preserve the main-95 handwritten enqueue_account_storage_cleanup() function and user_account_storage_cleanup trigger; generated snapshots do not describe these SQL objects.
Target-role dispatch resolves and validates the approved runnable provider once, before building context. After resume reads and prompt/model preparation, the existing memory-version guard checks complete Knowledge versions and the full selected owned resume in a short owner-first transaction. A shared resume row lock prevents edits while these versions are checked; locks release before provider I/O. Changes during preparation reject before transmission; the same guard rejects responses changed during HTTP. The full resume data is checked even when only its first 10000 Markdown characters are sent. No provider re-resolution or new lock/storage framework is involved.
Role preflight checks the captured connection’s availability in the same owner/Knowledge/resume transaction, after validating context. Explicit Off/delete or managed-AI removal committed before this preflight authorization prevents dispatch. A configuration retarget followed by a successful test still uses the previously approved captured configuration; a retarget left untested is unavailable. Locks end before provider I/O. Changes after committed preflight can race with transmission; the post-response check rejects stale results but cannot recall data already sent.
Role suggestions disable SDK automatic retries for this revocable prompt only. A manual retry resolves the approved recipient and checks the selected context again. Other generation callers keep their existing retry policy. A failed connection attempt or retryable HTTP response never silently repeats this prompt after its context check.
Reply-log snapshot contract
Read the full application withGET /api/openapi/applications/{id} and retain its opaque replyVersion while the user reviews it. Send that value as expectedVersion to POST /api/openapi/applications/{id}/replies. The server checks the complete stored snapshot under the application lock. A 409 CONFLICT requires a fresh read and user review; never automatically retry with a newer token. The token is not a timestamp or a value clients derive. Full application responses, including mutation results, provide it; summary rows do not. Older servers may omit it, so clients keep Copy available and disable explicit logging.
Explicit replies store followedUp.kind: "reply"; editable note text never determines contact kind. Legacy follow-up markers remain valid.