cesnimda
11 KiB
P6-02 — Existing AI interaction review
This is a source/caller and regression-contract review of existing approved tools, not model quality evaluation. The Phase3 AI architecture, AI principles, original contracts and P5-13 return conditions remain authoritative. No model/provider call, qualification, newly designed CV review UI or activation is performed.
Existing presentation and mutation boundaries
| Entry / purpose and actual owner | Source/pending/error/review behavior found in source | Preservation/test evidence and remaining qualification |
|---|---|---|
| Job analysis, match and timeline — ApplicationIntelligence | Read-only backend answers load by job/category, with PageSection/ViewStateNotice and retained text/chip/data context. Mount reads existing data; it does not start AI. Internal useIntelligence suppresses results after unmount, but error shell currently declares no Retry action |
application-intelligence tests, job-sections; W02/W03/E09. Source/revision/model provenance is limited to returned contract; cannot certify claim truth from a chip |
| Strategy snapshot — ApplicationWorkflowAssist | Explicit Generate queues an operation; stable operationID, progress stage/percent,1000ms polling while nonterminal, existing allowed Cancel/Retry and terminal toast dedup. Existing plan remains distinct from queue/result. No auto-generation on navigation | workflow-assist tests, StrategySnapshotOperation backend tests; W03/W12/S04. Enabled-worker F20 and actual model/delivery quality remain independently unqualified |
| Status/workflow suggestion — ApplicationWorkflowAssist | Existing status identity and rationale/signal are displayed. Suggested status is not a newly created enum/pipeline or hiring probability; any write remains existing explicit handler | workflow trust, JobApplicationsWorkflowSignals/ApplicationIntelligence backend tests; W02/W03. No global change to statuses for visual consistency |
| Tailoring/source CV — ApplicationAssets | Existing job-attached CV relation and selected source are shown; create/copy/tailor uses current owner API. This changes/selects a variant; it does not silently promote CV changes to canonical profile | application-assets, native preservation; W05/C07/E09. Broader live tailoring quality remains unverified |
Cover letter — local CoverLetterAiAssistant in ApplicationAssets |
Purpose/tone/action/document language/additional instructions and attached CV name; busy Generate; persistent local error with explicit Retry; Current/Suggestion side-by-side (stacked narrow). Apply puts proposal into local letter draft, Reject clears proposal; separate Save/history owner persists. No send implied | application-assets, job-sections, ApplicationAssets/AiWorkspace backend tests; W06/M04. Draft source freshness and delayed generation need bounded follow-up below; mocked responses do not establish model factuality |
Interview preparation — local InterviewAiAssistant in InterviewPrep |
Attached CV source, explicit focus/category/language/instructions, busy Generate, persistent error and user-initiated rerun. Add to prep explicitly persists one AI-sourced entry; Reject clears proposal. Answer edits remain local until Save and physical interview schedule is not inferred | interview-prep, InterviewAiContext/InterviewPrepPersistence backend tests; W09/W10. Focus freshness/review and actual generated personal facts require independent qualification |
| Career extraction/import — CareerProfilePage | Existing upload/process/operation/diff/low-confidence selections/accept/discard/retry/cancel. Canonical profile changes require explicit selected review, with manual editing available. It is a distinct extraction destination, not CV proposal Apply | profile-page tests, ProfileCvController/CvProfileDiffService tests, native keyboard picker in responsive audit; C01–C03/E09. OCR/corpus/model accuracy/full human review remain open; do not simplify away low-confidence controls |
| Existing CV section assistant — AiSectionAssistant | Purpose/action/document language, editable proposed text, Current, fact-check reminder, Apply/Reject/Close. Local requestID invalidates proposal on text/target/language/entitlement change and ignores stale response. Error uses shared toast; source/apply delegates to one editor owner | ai-section-assistant, CV editor tests, CV grounding/readiness; C11/C10/S04. Production withheld, not successful review UI completion |
Existing whole-document CV assistant — DocumentAiTab/editor callbacks |
Block IDs/structure/source are retained; invalid/missing/duplicate blocks cannot apply. Editor flushes current queued save, verifies epoch/revision, sends expectedVersion, applies server-issued proposalToken, and handles conflict. RequestID/target content invalidation retains account/source boundaries | CV editor/CvAiGrounding/CvAiReadiness/CvRevisionReadiness tests, real-HTTP withholding contract; C10/C11/E09. Deterministic guards cannot qualify factual generated output or replace absent P5-13 redesigned review |
| Privacy, entitlement and metering — existing accountPlan/AiPrivacySettingsCard/AiUsageCard/API | Existing server-owned privacy/ownership/entitlement and usage limits govern action; unavailable/locked state must preserve manual alternatives. Client canUseAi is presentation, not authorization | AiPrivacyPolicy/AiUsageMeter/AccountPlans/API authorization suites, AI readiness; S04/E09. Actual provider/checkout config is not proven by these tests |
Consistency findings and bounded future reproduction
Common established pattern is explicit contextual invocation→bounded pending/error→review→deliberate domain action. Source/destination vocabulary is meaningful: Apply to CV can autosave via the established variant owner; Use in letter creates a local draft with separate Save; Add to prep creates a reviewed persisted item; accepting extraction changes selected canonical fields; none authorizes email send. These differences are intentional and must survive any presentation reuse.
The existing letter assistant has an explicit Retry button in its inline error; interview reruns through Generate; CV errors use existing toast and Generate. Read-only intelligence error presentation has no local Retry. This is a review finding for remaining domain migration, not evidence that a failed request erased original content. Reuse existing shared alert/state infrastructure only after retaining domain recovery behavior; no second notification/review framework.
AI-SOURCE-01 — pending letter context: source inspection finds no requestID/original-text fingerprint invalidation inside the local letter generator. To reproduce safely later, intercept/delay a synthetic generation response, change the editable manual letter while pending, settle the response and inspect Current/Suggestion/Apply. Expected acceptance is a review explicitly tied to current source or invalidation requiring regenerate; preserve the edited original and no-write Reject. Current Apply sets only a local draft and explicit Save remains, so do not misreport this source concern as proved automatic persisted data loss. No native failure is claimed here.
AI-SOURCE-02 — pending interview focus: generation snapshots selected focus into context, while the selected focus control remains editable and acceptance uses the current selected category/title. A delayed synthetic response with focus change is the required reproduction; verify stale review/category behavior and user recognition before correcting the narrow owner. No request/acceptance/backend policy is broadened for style. These existing controls belong to the Phase7 remaining application workflow review unless a fresh proof demonstrates a severe present defect.
Regeneration, source edits, cancelled operation, late response, missing/foreign record, account expiry/change and invalid proposal deserve the existing boundary assertions. A backend guard passing does not imply every legacy UI has the same local conflict affordance. Do not merge disparate Apply callbacks, request payloads, revisions or save/send semantics to reduce duplicate JSX.
P5-13 remains withheld
PRES-C11/P5-13 is DEFERRED / UNQUALIFIED / REVIEW UI NOT IMPLEMENTED. The actual configured-target frozen32-case/366-call qualification has0 completed calls and0 independent human semantic reviews. Candidate runtime/model/hardware information is not measured approved identity. Production registration withholds CvAiSuggestionService; generation/Apply return established503 cv_ai_unavailable before provider request/content/history write. Existing unrelated AI paths remain.
Future qualified P5-13 UI should reuse existing labelled source/destination/current/proposed comparison, non-color review semantics, busy/progress/error/Retry, explicit Accept/Reject/edited proposal and stale-context conflicts. Its source/block/proposal token/owner/entitlement/grounding validators remain authoritative and domain-specific. Manual workflow remains reachable when unavailable. This review neither designs/activates the withheld capability nor runs model qualification.
Fresh applicable frontend/backend/sidecar/real-HTTP guard results must come from the P6-05 manifest. Human AI semantics, reader interpretation, live model quality/provider operation remain NOT VERIFIED. Keep all qualification/deferred requirements open, and keep Production NO-GO.