[DEV-2203] App AI tools, cited research and media guides

[tonytlwu]

JIRA ticket

DEV-2203, DEV-2204, DEV-2205

Summary

Document three app AI workflows: validated tool calls and confirmed authorized changes, web research with inline sources, and PDF/photo extraction with editable structured results. Extend the JavaScript reference for the additive completion lifecycle in API PR #8804.

Changes

• Complete browser examples, model-specific payloads, bounded workflows and explicit failure/permission boundaries.
• Private files are fetched under Media access rules and only authorized bytes are sent to AI. No provider keys or authenticated Media URLs enter the model payload.
• Correct unsafe custom-rule request interpretation: validate the route-supplied filter for constrained query reads and operation payload plus stored ownership for writes. Preserve legacy query; never treat arbitrary body fields as an applied filter. These contexts depend on API PR #8804 and fail closed when absent.
• Clarify stable protected owner keys: verify identity or provision trusted accounts before granting preassigned private records; public signup alone does not prove control of a claimed email.
• Preserve one AI namespace, regenerate discovery indexes and add reciprocal guide links.

Verification

• 236 documentation unit tests passed, including execution of the actual fenced examples, unsafe citations/Markdown, typed results, permissions, incomplete/refused output and explicit retry.
• Strict indexes, Jekyll build, raw Markdown siblings and whitespace checks passed. Jekyll retains existing warnings in unrelated examples.
• Corrected actual fenced security rule: 34 live HTTP cases pass, including the reproduced no-where cross-owner read attack, disguised write payloads, proper owner reads/updates, wrong login, anonymous, direct-read denial and atomic mixed-owner denial. API also passes 27 actual-model filter/payload tests. Earlier disposable fixture rules were replaced.
• Local real app-token/provider recipe checks: tool read/replay, declined change leaves Open, accepted fixed assignment readback, hosted sources and private media transport. These recipe checks are separate from Studio-generated app completion.
• Supplied encrypted PDF: owner bytes match original; other token 403, unauthenticated 401, unsigned backing-object access denied. Denied/expired browser retrieval produced zero AI calls. Actual extraction returned invoice #20130304, total 39.60, null missing date and null currency when only an ambiguous $ is printed.
• Supplied golf photo produced validated editable visible details. Real hosted search returned completed search items and source annotations.

Presentation follow-up

• Broaden the existing presentation guide/cross-references to interactive generated prose, including research. Preserve plain-text requirements, literal code, structured results and editable fields. No new page, API, runtime helper or mandatory renderer.
• Clarify streaming and formatting as independent choices. The buffered grounding example demonstrates citation handling; link progressive text to final native-result citation processing. Keep provisional text unverified and preserve original annotation indices before formatting a display copy.
• Studio preserves applicable display requirements through research and executor context, and keeps safety conditions from becoming capability bans. Provider details remain in fetched documentation.
• Local app58 now uses pinned lazy parser/sanitizer resources through normal Fliplet dependencies and Vue reactive rendering. Raw output/history and original citation offsets stay intact. 23 actual-source fixtures and independent review pass, including hostile content, split streamed syntax, citation preservation, renderer failure/recovery and final/history parity.
• Live app58 search and contextual follow-up: semantic headings/lists/emphasis visible while Drafting, final answers and history remain formatted, returned source opens PubMed. Desktop/mobile screenshots inspected with no horizontal overflow. This is a targeted repair, not evidence of first-pass generation consistency.
• Studio: 107 focused tests, scorer/evidence checks including 38 evidence cases, canonical lint zero errors/warnings. Docs: 236 tests, strict indexes, Jekyll, raw siblings and whitespace pass. No application dependencies changed.
• One neutral fresh build, app69, initially chose buffered output and literal Markdown despite complete writer retrieval. Actual task allowed text OR sanitized Markdown; the research summary overemphasized the buffered example. The parent independently noticed readability and repaired it without coaching: real final/follow-up output now shows headings, emphasis and linked sources. Version 1 is saved. Requests remain buffered, so progressive answer display still fails; first-pass presentation is not a PASS. The in-flight build does not validate the later summary/docs refinement. Completed trace: 9/9 retrieval/native-call/model checks with all nine executor lanes captured; that score does not measure formatting or progressive display. No repeated builds or coaching used to manufacture a pass.

Deployment instructions

Keep this PR in preview while testing Studio. The terminal-result reference and custom-rule filter/payload examples depend on API PR #8804 and its published SDK asset. Deploy/verify the API and SDK before publishing these contract-dependent docs, then refresh Studio's production catalog from published docs. No production deployment is requested by this draft.

Author concerns

Final preview raw Markdown/index retrieval and rendered recipe desktop/mobile checks passed; the preview content matches reviewed source. Docs unit CI passes at 4ac3b8248714a367053ce6310e1401e3aceeafa4; Cloudflare preview also passes. Fresh generated-app outcomes remain separate: PDF extraction/editing works, but Save/private-original retention and tax/warning review remain gaps. Targeted local action app repairs now pass real provider read/replay, declined unchanged and accepted Status-only updates plus cross-owner denial. Research now passes progressive/final/source/follow-up UI checks after the shared nested-template loader repair in API PR #8804. These repaired-app outcomes are not evidence of consistent first-pass builder success. Safe Markdown is now repaired and verified in local app58; the unchanged chatbot control still renders literal Markdown. This is not a generic agent framework, exactly-once action service or offline AI feature.

[cloudflare-workers-and-pages]

Deploying fliplet-cli with    Cloudflare Pages

Latest commit:	4ac3b82
Status:	✅  Deploy successful!
Preview URL:	https://1ff80bcf.fliplet-cli.pages.dev
Branch Preview URL:	https://feature-dev-2203-app-ai-capa.fliplet-cli.pages.dev

View logs