StoryFlow Public Progress
This is a sanitized public status summary, not the protected detailed milestone history. FlowTurbine parity and acceptance retains all 41 requirements, their archive anchors and exact acceptance scenarios, and documents the candidate workflows and actual source-bound shortcuts.
D-202 bounded authenticated continuation — browser phase complete
The D-198 result remains immutable and BLOCKED at its dated closure. D-202 permits a new, narrowly bounded acceptance continuation; it does not retroactively turn the unavailable D-198 browser transport into a pass or waive any of the 41 exact scenarios.
A separately owned NI10 headless Chrome completed genuine pxl authentication through Authelia and opened the public application. The verified host-resolver mappings for app, auth and wiki point to public IP 178.18.250.170. Captured app/auth responses use TLS 1.3, authenticated API/snapshot responses return HTTP200, and wss://stfl.loca.zone/ws upgrades with HTTP101. The browser’s loopback CDP endpoint is control transport only, not the application route; TLS verification is not bypassed.
The first browser-tooling probe is explicitly quarantined. It reported a renderer in an unrelated NI0 browser and did not establish the required launch arguments or ownership. Its observed sign-in/TLS is not counted as D-202 acceptance. That process was not killed or reniced; the separately owned browser repeated the genuine sign-in and public transport observations.
Main acquired the NI10 kernel WRITE heavy lease and invoked the sole authorized native continuation once, retaining the failed final one-reference/unlocked-seed/8-second request’s provider/model/settings. It exited 2 with checkpoint_clone_changed before confirmation, reservation or provider activity: isolated database/WAL matched but SHM differed, with cause unattributed. Provider POSTs and native-fetch invocations were zero. This is an admission refusal, not success or another HTTP403. The allowance is consumed; no replay, checkpoint repair, rotation, switch or other generation is authorized. The bounded browser/local-media phase is finished; final verification/publication remain Main’s responsibility. D-198 images/videos/403 and costs remain dated history below.
Real credential lifecycle remains unaccepted. Source inspection of the deployed Credentials surface and native-auth adapter finds only shared global OMP storage, not a supported disposable app-scoped path that preserves existing authentication bytes. Opening/refreshing that store can mutate it. Viewing metadata or opening and cancelling a consent dialog must not be called add/test/update/remove success; the same limitation governs credential consent and the unchanged Credentials manager acceptance.
Frozen browser observations — not complete parity
The following observations precede the bounded cleanup recorded below. Cleanup does not erase their receipts or make the complete parity contract pass.
| Surface | Observed in the public authenticated browser | Remaining boundary |
|---|---|---|
| Authoring and imports | Ordered CSV/JSON/TXT drafts, variables/literal braces and row-specific errors; actual TXT drop; edit/reorder/duplicate/draft reset. Keyboard JSON-record import preserves seeds 101/202, variable precedence and separately populated direct/settings references. A newly promoted draft with nonempty start/parent links was deleted, exactly restored by Undo and re-deleted. | CSV BOM remains unexercised; the first TXT drop was interrupted by disclosure closing. Draft reset is not completed-work reset. The restored draft’s explicit refs array was empty and it had never completed; tombstone/undo/log residue remains. |
| Library and mapping | Picker/file-drop bytes, tags and role mappings survive reload. Native Library payloads target start/end/reference fields; drag and @tag resolve the same asset, Cancel is exact and reuse is canonical. Keyboard clear/share/override/append/Save and direct end-frame mapping work. A TXT Library drop is visibly rejected without asset/job changes. | No fresh frame/reference provider submission. Duplicate-tag assignment fails before ambiguity; a separate ambiguous-resolution state was not exercised. Gallery-origin role drops and every keyboard equivalent remain open. Failed probes remain preserved. |
| Batch export/import | Literal Chrome-downloaded mapped-batch JSON parsed and imported into the other owned production: three ordered shots and two canonical remapped assets, matching prompt/params/model/seed/lock/settings/reference roles/ingredients. Embedded image bytes/digests match; no credential-field keys were found. Invalid .json picker upload was visibly rejected without batch/shot/asset/job changes. | The tested invalid file was malformed JSON, not a mixed-valid/invalid transaction rollback. Export still omits some production settings and is not a full backup or universal secret-DLP guarantee. The separate historical-video inclusion superset remains explicitly different from a literal original endpoint export and creates three unique assets, not five. |
| Queue and preflight | Presets 1/2/4 and custom 7/5 persisted; cap 5 changed while held. Four held variants and two ordered held jobs stopped before dispatch. Batch search and ordered all-eligible-draft preview show exact settings, unknown costs and explicit capability/seed/frame/reference blockers; both Confirm actions are disabled and keyboard Cancel preserves jobs/shots. | Saved limits are not measured in-flight bounds. Attempt ordinal 1 has an empty attempts array. No Run All dispatch, mixed running/retrying completion or selected-item Retry-to-terminal is established. |
| Logs | Three real upload records were filtered, copied with exact readback and downloaded as matching 174-byte TXT. Follow resumed at the bottom of overflowing content. Clear Cancel preserved ten owned logs; confirmation/reload cleared only those logs, leaving jobs and other-production logs unchanged. | This is authoring/upload activity, not generation-success/failure correlation or universal secret-redaction proof. New-entry behavior while browsing remains unexercised. |
| Keyboard and narrow/wide surfaces | All displayed Linux shortcuts, typing suppression, nested Escape and exact-opener return; keyboard select/map/compare/import/edit flows; sampled 320px/390px/1512px geometry, contrast 8.68–15.57, reduced motion and visible focus. | Not every mutable flow, exhaustive accessibility or stress coverage. Native Queue still has a narrow-error layout defect; no-overflow is insufficient. Captions remain unavailable. |
| Gallery, promotion and lineage | Select All reports 6, Clear disables selection actions, and keyboard selection compares four compatible images: two historical originals plus local 2K/4K derivatives. Corrected keyboard clearing/restoring slot 4 preserves the first three. A two-ID native selection drop populates slots 1/2; the earlier four-ID drop to slot 4 failed. Historical-video selection survives reload. Separate lineage examples cover converging video parents and original→2K→4K parent/child/Back navigation after reload; a promoted draft preserves its image start/parent. | Four compatible images are not four newly generated native variants. Source prompts are unavailable with explicit current-shot fallback; original structured dimensions are unavailable despite real rendering. No compared-winner/sibling acceptance or fresh promoted video. The separate legacy snapshot failure remains. |
| Media downloads and naming | Real local Lanczos outputs are 1367Ă—2048 and 2734Ă—4096 PNGs, fully decoded. Historical video plays in the web browser and fully decodes as 192 H264 frames, 8 seconds, 1280Ă—720, no audio. Exact unprefixed naming saves/reloads; actual ZIP member names, CRC/SHA, asset mapping and safe unique paths are verified. Traversal naming is visibly rejected without mutation. | Local resize is not AI enhancement and historical playback is not new/text-only generation. Colliding derivative names are disambiguated by asset ID, but the 4K name changes with selection: collision avoidance is not selection-independent stability or multiple-variant proof. |
| Download scope escape | The 4K helper timed out/detached before the blob download; Chrome recorded its default Downloads destination outside permitted evidence roots. The file was not touched or removed. A later scoped download reused the already-created 4K derivative without another resize. | An explicit unresolved scope violation/residue, not repaired by the successful scoped download. |
| Installed sidepanel and resource isolation | Actual Chrome SIDE_PANEL context and 360Ă—412 surface, public-IP/TLS1.3 REST and snapshot HTTP200, selected-production reload and existing-image rendering. No Flow tab or granted/requested Flow permission. Unrelated public wiki-origin panel/chunk fetches and iframe access were blocked by Chrome. | Background WSS handshake returned HTTP302 and the panel remains live-updates disconnected; web-face HTTP101 is not extension success. Queue error text collapses to a 42.27px column despite no overflow. A loading video is not playback; no panel generation was run. |
| Credentials | The explicitly chosen provider screen, shared-store warning and actual consent Cancel were exercised; the native lifecycle was not confirmed. | No supported disposable app-scoped add/test/update/remove path exists. An initial whole-row authentication hash change is unattributed; later credential-data stability with metadata-only change does not prove initial full equality. |
Bounded cleanup used the supported authenticated public batch-delete API for all 11 owned batches. All 23 newly added shots are soft-deleted and both owned productions now expose zero active batches/shots. Six stopped jobs retain empty attempt arrays. Six remaining owned Library deletions returned HTTP409 because tombstoned references, requests or derived-parent edges still protect them. Productions, historical Gallery video, two derivatives, media bytes, tombstones, undo records and attributable logs/events remain: this is not zero-residue cleanup. The outside-root download is outside permitted cleanup scope.
Original StoryFlow application rows remain present and unchanged across all 13 tables, including 34,744 events, 11 assets, 5 batches, 17 jobs, 9 productions and 16 shots. Actual hashes of all 11 original media match their unchanged recorded digests; three legacy zero-byte-count metadata defects are separately retained, not repaired. All 501 pinned original/rollback/anchor/result artifacts and current D-198 project-routing bytes remain unchanged. This does not resolve the earlier global-authentication hash inequality. Earlier public/private wiki route observations precede this source freeze. Main owns final checks, publication, rendered-route/private-split verification and closure; the receipt index preserves the boundaries.
D-198 — deployed locally verified; mission BLOCKED
The canonical face and gateway source for the StoryFlow application are deployed at source commit fa9ab38a32a4ca805ca882d6480d3173bffc1c7d. This is actual owned-service/static-byte deployment, not required external Chrome/realpxl UI or authenticated public WS101 proof. That acceptance remains unavailable, so the mission verdict is BLOCKED. Native evidence is unchanged: four images, three videos and one terminal HTTP403 across eight POSTs, with no further native work. Canonical build and controlled duration-guard proofs remain separate from those earlier paid runs.
No row is declared complete. Earlier local receipts cover the versions they exercised; they do not automatically certify later repairs. The frozen archive remains read-only reference material.
Deployment and per-item delivery — 2026-09-28
The deployment receipt records gateway PID 2207413 running actual Bun 1.4.2 with inherited NI10. Both /health and /api/health return 200. All 13 served static-file SHA hashes match the canonical build. Owner-host loopback verification proves those exact responses/bytes, not public browser rendering, UI interactions or authenticated WS101.
Every prior live row remained unchanged: 11 assets, 17 jobs, 9 productions, 16 shots, zero ledger entries and all 32843 prior events. Four new startup events bring the event count to 32847. A verified WAL-aware consistent snapshot, raw DB/WAL/SHM and previous static release were privately preserved for rollback; no database or backup contents are published here.
All 41 per-item commits/receipts were completed and independently verified, ending at the deployed commit above: 69 source paths, 37 source patches and evidence-only rows 20/21/22/24. Delivery accounting is not full acceptance. All 41 exact required scenarios remain intact, with unobserved or incomplete outcomes still identified below and in the acceptance table.
Evidence observed locally
The preserved isolated gateway receipts exercised real HTTP, SQLite, WebSocket server echo and FFmpeg, without changing the live database or generating cloud media. Observations include genuine image-file ingest, production-scope rejection, CSV/JSON materialization and atomic invalid-import rejection, semantic batch roundtrip, reorder/delete/undo, quote cancellation/idempotency/staleness, held/start work, persisted queue settings, selected retry/restart, local winner persistence and correlated logs.
Real local concurrency observations include Slow 1, Balanced 2, Fast 4 and custom 3, including a queued change from 1 to 3. Actual image resize outputs were 2048Ă—1152 and 4096Ă—2304; those are FFmpeg Lanczos results, not AI enhancement. Multi-asset ZIP proof covered production folders and unchanged original bytes. Local assembly outputs are not generative-video proof or browser four-up comparison proof.
A controlled persisted-429 HTTP scenario rejected manual retry and new confirmation without creating jobs. This was not an actual provider 429. A transport-helper check produced same-origin REST/assets and secure WebSocket URLs; isolated WebSocket echo is not remote authenticated WS101 acceptance.
The corrected caller runtime smoke also passed through actual isolated HTTP/SQLite/FFmpeg: Cancel created zero jobs; a blocked quote never invoked approval; held confirmation made no execution attempt before Start; replay returned the same job; and the resulting local MP4, download and parent lineage were observed. Approval was programmatic, not a human or remote-browser action. The previous probe’s mistaken assumption about the queued attempt ordinal remains preserved; the corrected check measures execution history rather than treating ordinal 1 as an execution.
Genuine native results and preserved interruptions
Four genuine 1376×768 JPEGs were generated through existing OMP authentication on google-antigravity/gemini-3.1-flash-image, using count-one then count-three paths across isolated runs. The first image was retained/re-imported into Library without regeneration; its original generation provenance stays in the prior receipt, not a fabricated restored job. Image costs remain unknown and seed_applied is false. An FFmpeg-composed four-up board is media evidence, not the app’s compare controls or a browser screenshot. API winner/promotion and isolated reopen persistence were observed.
Three genuine openrouter/google/veo-3.1 videos completed: two frame-input clips at 4 seconds each and one two-reference clip at 8 seconds. Each is H.264 MP4, 1280Ă—720 at 24 fps, with no audio stream; complete FFmpeg decode passed for all streams. The 4-second clips contain 96 frames each; the 8-second clip contains 192. The two frame runs submitted identical ordered private start/end bytes and locked seed 1980928. Different output digests do not establish deterministic output. Contact-sheet observations show the two frame clips transitioning toward the authored end frame and the reference clip changing camera viewpoint; no browser playback or UI acceptance follows.
The final one-reference scenario patched the same Shot, removed one reference, unlocked the seed and persisted/submitted seed 1797383276. It received HTTP403 on its sole attempt, produced no output and remained failed with no retry timer, account/provider/model change or further generation. The precise upstream refusal reason is not established. Mapping, seed, jobs and assets survived shutdown/reopen without native redispatch. No actual structured 429 occurred.
The successful videos report 80 + 80 + 160 cents, converted from provider USD and rounded to integer cents with an explicit nearest-cent basis. This known-video subset is 320 cents, not an invoice or mission total: all image charges and the failed request’s charge remain unknown.
Earlier failures remain history. The first journey exposed nested PATCH defaults resetting aspect on a count-only update and discarding frames on a references-only update. The actual-HTTP Shot/Batch repair gate passed 6 tests/42 expectations in 4.08 seconds; subsequent native continuation also preserved settings and rejected mixed frame/reference inputs. Two earlier operator approval timeouts created no jobs and were not provider failures; separately approved continuation quotes then generated the remaining outputs. Parent/operator confirmations are not the authorized browser approval journey.
A later source preflight guard rejects documented Veo reference-input, 1080p or 4K requests unless duration is 8 seconds, without silently rewriting the request. Its separate actual-HTTP smoke now passes 12 quote cases: settings references, direct refs, ingredients and 1080p each block 4/6 seconds and admit 8 seconds. Quoted durations remain unchanged, total cost stays null/unknown, and the smoke creates zero jobs/provider calls/unexpected network requests; only capability metadata is controlled. The permanent HTTP regression separately observes invalid confirmation HTTP400 and one valid 8-second held job, not generation. The final focused gate passes 55 tests, 0 failures, 359 expectations across four files in 26.95 seconds. Earlier 52-test and 6-test receipts remain historical snapshots; native paid proof precedes this guard.
Frontend build recovery and subsequent deployment
The owner traced the earlier frontend failure to Vite 6.1 ignoring WXT hook filters. The preserved successful checkpoint used WXT 0.21.4, Vite 8.3.1, Svelte 5.57.1, SvelteKit 2.70.3 and Svelte plugin 7.3.1. Earlier missing-entrypoint failure and two timeouts remain preserved.
That checkpoint’s panel build passed in 12.8 seconds (WXT finished in 18.7 seconds); staged face passed in 17.18 seconds with 184 SSR and 279 client modules. The generated MV3 manifest has only sidePanel/storage permissions, the StoryFlow host permission and no web_accessible_resources declaration. This is generated-file evidence, not an installed-extension test; SSR compilation does not mean the static face is deployed SSR.
The audit found a normative dependency violation: the earlier successful lock used daisyUI 5.0.0-beta.8 and TypeScript 7.0.2 despite sum.yml requiring daisyUI 5.7 and TypeScript 5.9, not TS7. That receipt and its artifacts remain preserved under frontend-build-history/pre-canonical-deps; the violation is not rewritten as a contract change.
The canonical daisyUI 5.7.0/TypeScript 5.9.3 locked-install and both rebuilds pass. WXT panel build took 4.193 seconds (finished in 5.628); face took 21.58 seconds with 184 SSR and 279 client modules. The generated manifest retains only sidePanel/storage permissions, the StoryFlow host and no web_accessible_resources. Actual compiler-API smoke passed with zero parse diagnostics; no full-project typecheck was run. The canonical face/source were subsequently deployed with the separate local proof above; neither build nor deployment establishes extension installation or external-browser acceptance.
Retained warnings include real captions being unavailable, future native-config-loader use of __dirname and deprecated optimizeDeps.esbuildOptions. No fabricated captions or warning suppression is claimed. Accessibility, responsive layout, shortcuts and authorized external-browser acceptance remain unestablished.
Failures are still failures
| Preserved observation | Meaning at this checkpoint |
|---|---|
| Initial focused gate: 160 passed, 1 failed, 303 expectations across 7 files; 15.90 seconds | Failed overall. The transient-item/real-sibling case timed out at 10 seconds and the runner reported two dangling processes. This does not justify assuming CPU contention or merely increasing the timeout. |
| Subsequent queue diagnostic: passed in 8745 ms | Controlled transient first failure plus real isolated SQLite/FFmpeg sibling and unchanged-request retry succeeded. This is specific local evidence, not a replacement green integrated gate or cloud retry proof. |
| Security-before: provider acceptance followed by local ENOENT allowed manual retry and two controlled generation calls | Reproduced a paid-submission resubmission safety defect. The matching security-after now leaves the item blocked as accepted_media_interrupted, refuses retry and observes one controlled generation call. No cloud call occurred. |
| Security-before: login near session expiry lost its flow; poll returned 404 and cancel reported missing flow | The matching security-after now observes poll 200 and cancel 200 after the original session expiry, preserving flow ownership. This uses controlled login dependencies and isolated native storage, not real shared-store lifecycle acceptance. |
| Subsequent focused gate: 39 passed, 0 failed, 272 expectations across 4 files; 14.46 seconds | A passing isolated gate for its recorded snapshot, not erasure of earlier failures or coverage of later audit defects. |
| Fractional-cost reproduction: controlled USD 0.12345 completion | Before: integer-cents validation raised an unhandled rejection and left the durable job running; the failed probe did not serialize submission counters. After: one controlled submission completes with one asset and durable done state, storing 12 cents with an explicit nearest-cent rounding basis. |
| Image-input receipt reproduction | Before: changed image bytes were submitted with the original digest recorded. After: changed input fails as asset_changed before adapter submission, with zero controlled submissions and zero assets. Original source files remain unchanged. |
| Final focused backend gate: 52 passed, 0 failed, 287 expectations across 4 files; 15.19 seconds | Matching actual SDK/adapter/runner/queue/owned-SQLite/ffprobe media probe also passed with zero unhandled rejections. This establishes those controlled repair paths, not genuine cloud or authorized remote Chrome acceptance. |
The original failure receipts remain failures even when a subsequent repair succeeds. Final observations must be reconciled here before publication of a release claim.
Capability and authorization limits
Metadata alone does not prove entitlement. Preserved native wire receipts now additionally prove exact frame/reference submissions and the seven successful outputs above. They do not prove text-only video, every model/aspect/count option, seed determinism, arbitrary MIME/size limits, successful final one-reference generation or any browser workflow. Plain JSON prompt-record import also omits some current video settings, and batch export is not a complete production backup; see candidate workflow limitations.
The read-only OMP broker-status probe reported not_configured at its recorded checkpoint; it is not current credential inventory and does not contradict later existing-auth media dispatch. See native credential consent and side effects. Four image and three video successes do not establish real shared-store add/test/update/remove or the remote Credentials UI. The HTTP403 remains an unresolved provider refusal, not a reason to change accounts or retry.
The complete external-capability and archive reports have been recovered and normalized; original truncated wrappers remain preserved. Report recovery improves evidence completeness, not runtime/browser acceptance. No raw report or private evidence payload is copied into this public summary.
D-198 external browser proof — historical BLOCKED
Authorized non-loca, non-Mac Chrome/realpxl execution transport remains unavailable. Required acceptance must exercise the public application through an actual authorized sign-in, including UI interactions, responsive/keyboard behavior, reload persistence, same-origin API/assets and authenticated WebSocket transport.
Local or loopback Chrome, forged authentication headers, unauthenticated HTTP checks and source screenshots are not substitutes. Genuine media/API observations and actual owned-service/static-byte deployment are recorded separately; the required authorized remote-browser UI/WS101 proof is still BLOCKED. Real credential-UI lifecycle, text-only native video and successful unlocked/reference-removal output remain incomplete. Main owns final publication and mission reconciliation.
That paragraph records the D-198 authorization boundary, not D-202’s current permission. The bounded D-202 continuation above now admits the separately owned loca Chrome through public vhosts and records genuine pxl/TLS/WS101; it does not yet establish all UI workflows or erase the original BLOCKED receipt.
Public summary and protected history
Detailed milestone history was intentionally moved behind sign-in in D-164 and remains reachable only through the existing private navigation. It is not copied here. Raw evidence, SQLite/WAL files, credential values, OAuth URLs/callbacks, private search entries and private assets do not belong in this public summary. The public Quartz private-content exclusion must remain intact.