AlloFlow Document Remediation

Convert the blanks you accepted from detect_form_fields into real, labelled form inputs in the HTML. Deterministic, no API key. Pass the production pipeline's native object map keyed by candidate id, for example `{ "f0": { "label": "Student name" } }`; anything omitted is left as-is.

Run the same two-engine accessibility audit the remediation pipeline uses, on a local .html file: the AI content rubric (score, per-issue WCAG mapping) plus the axe-core automated engine. Title II is web-first — use this to evidence-check web page exports, LMS content, or any saved HTML without remediating it. Local FILE input only; this tool never fetches URLs, so nothing leaves the machine except the audited HTML going to the Gemini API under your key. Writes no files. Typically 1-2 minutes.

Audit an accessible HTML file with axe-core AND IBM Equal Access, then report where they DISAGREE. Both run locally in the browser with no model and no API key. This matters because a single engine is a single opinion: on this project axe reported 100/0 on a document veraPDF then failed, and reported no contrast findings on text at roughly 1.6:1. Two engines that disagree give a short, concrete list a human can check. Use it before telling anyone a document is clean.

Report the heading hierarchy of an accessible HTML document (counts per level and any SKIPPED levels, e.g. h2 followed by h4, which is a WCAG 1.3.1 failure) and produce a plain-text rendering. Deterministic, no API key. Plain text is itself a legitimate accessible format for readers who want no markup at all, and the heading check is the cheapest way to catch a structural mistake before tagging a PDF.

Run AlloFlow's vision pass over the data-URI images embedded in an accessible HTML file: describes each one, classifies equations and charts, marks recurring chrome (logos, letterheads) decorative, and groups near-identical images so a repeated logo costs one vision call rather than one per page. Writes alt text into the HTML. REQUIRES a Gemini key and sends image content to the API — this is the one HTML operation here that genuinely needs a model.

Scan an accessible HTML document for places that should become fillable form fields (blank lines, underscored runs, answer boxes) and report each with its label, kind, width and surrounding context, flagging any that have no label. Deterministic, no API key. Pair with apply_form_fields to turn the accepted ones into real labelled inputs. Unlabelled fields are a WCAG 4.1.2 failure, so `labelMissing` is the field to act on.

Convert an accessible HTML file (the -accessible.html a remediation produced) into an accessible Word (.docx) or OpenDocument (.odt) file, preserving heading structure, lists, and tables. Deterministic packaging: needs NO Gemini key and spends no quota, though it does fetch React and JSZip from public CDNs. Use when someone needs the remediated document in an editable Office format rather than a PDF. For ePub 3, DAISY 3 or Braille, use export_alt_format instead.

Convert an accessible HTML file into an ePub 3 ebook, a DAISY 3 full-text talking-book package, or an uncontracted (Grade 1) Braille BRF file for embossers and refreshable displays. Deterministic restructuring: needs NO Gemini key, spends no quota, and packages entirely offline (unlike export_accessible_office it does not need JSZip from a CDN, though React is still fetched). Use when a student needs an ebook, a DAISY reader, or hard-copy braille rather than a PDF. ePub output is checked by a built-in structural self-check whose findings are returned as `structuralErrors` — that is NOT epubcheck, so a book reported valid here can still be worth running through epubcheck. Braille is Grade 1 only (contracted UEB needs the liblouis plugin, which is in the app, not here) and reports how many characters had no braille equivalent and were dropped.

Pull the text out of a .docx, .pptx, .xlsx or .pdf deterministically — mammoth for Word (preserving heading styles and list structure as markdown), the spreadsheet converter producing markdown tables, pdf.js for PDF text layers. Needs NO Gemini key and sends nothing anywhere. Reports the extraction `method` and any `error`, so an empty result is distinguishable from a failed one. Use to inspect a document before deciding how to remediate it, or to check whether a PDF has a usable text layer at all.

Apply AlloFlow's deterministic contrast repair to an accessible HTML file: darkens or lightens foreground colours that fail WCAG AA against their detected background, then sanitises remaining style for WCAG. Pure colour math — needs NO Gemini key and spends no quota. IMPORTANT: the returned axe numbers do NOT verify contrast. axe-core does not reliably detect contrast in this harness (measured: zero findings on text at ~1.6:1), so `styleFixes` (the count of corrections actually applied) is the evidence, and a human or a contrast checker should confirm.

Produce AlloFlow's own Accessibility Conformance Report as HTML, the same report the app generates, from artifacts you already have: an axe audit and (optionally) a veraPDF verdict. Deterministic templating — needs NO Gemini key. Use this instead of hand-writing a report, so what an agent hands a user matches what the application produces and cannot drift from it.

Generate the same student/teacher resource-pack HTML as the normal AlloFlow app by calling its existing generateFullPackHTML pipeline export unchanged. The input JSON uses the app's native resource shape, so this is a thin adapter rather than a second renderer. Deterministic: needs NO Gemini key, account, or Worker. The JSON must contain `items` and may include `topic`, `isWorksheet`, `responses`, and `config`.

Run the AlloFlow accessibility audit on a local PDF, DOCX, PPTX, PNG/JPEG/WebP image, markdown/text/CSV file, or spreadsheet: overall score, per-severity issue list, scanned/searchable detection, page count, detected language. Sends document content to the Gemini API; core browser libraries are bundled locally. Writes no files. Office files are audited deterministically from extracted text (no Vision pass). Typically 1-3 minutes.

Start a BACKGROUND JOB that AUDITS every .pdf/.docx/.pptx in a folder (non-recursive, up to 200 files) and writes one triage scoreboard. This is the cheap first pass before remediation: an audit is 1-3 minutes and writes no document files, where a remediation is 5-30 minutes and spends far more quota, so run this to find out WHICH files are worth remediating instead of remediating a folder blind. Each document is sorted into a band that names the next action: scanned (image-only, needs OCR first), needs-work (<70), review (70-89), likely-ok (90+). Writes accessibility-audit-scoreboard.json and .csv (collision-safe, never overwrites); the CSV opens in Excel. Sends document content to the Gemini API. Poll remediation_job_status for per-file progress; remediation_job_result returns the scoreboard.

Start a BACKGROUND JOB that remediates every .pdf/.docx/.pptx in a folder (non-recursive, up to 60 files), one at a time, continuing past per-file failures. Returns a job_id immediately; poll remediation_job_status for per-file progress and fetch remediation_job_result for the per-file summaries. Same outputs and options as pdf_remediate, applied to each file.

Run the full AlloFlow remediation pipeline on a local PDF: audit, rebuild as accessible HTML, iterative AI fix passes to the target score, honesty-checked verification, and a tagged PDF export. Accepts .pdf, .docx, or .pptx. Writes <name>-accessible.html, <name>-tagged.pdf, and <name>-remediation-report.json next to the input (or to output_dir), never overwriting existing files (Office inputs skip the tagged-PDF export — the accessible HTML is the deliverable). Returns the distribution verdict, before/after scores, and every fidelity/honesty disclosure. Sends document content to the Gemini API. SYNCHRONOUS: blocks 5-30 minutes — if your client enforces a tool timeout, use pdf_remediate_start + remediation_job_status instead.

Start a BACKGROUND JOB that remediates only the documents a triage scoreboard put in the bands you name (default: needs-work). This closes the triage loop: pdf_batch_audit_start finds which documents need work, this fixes exactly those, so a folder that would be an overnight serial run becomes the handful that earned it. Point it at a scoreboard with scoreboard_path, or at a folder with dir_path to use that folder's newest scoreboard. Files listed in the scoreboard but no longer on disk are reported, not silently dropped. Same outputs, options, and resumability as pdf_batch_remediate_start.

Start the same remediation as pdf_remediate as a BACKGROUND JOB and return a job_id immediately. Jobs run one at a time in start order. Poll remediation_job_status (every 30-60s is plenty; runs take 5-30 minutes), then fetch remediation_job_result. Accepts .pdf, .docx, or .pptx. Writes <name>-accessible.html, <name>-tagged.pdf, and <name>-remediation-report.json next to the input (or to output_dir), never overwriting existing files (Office inputs skip the tagged-PDF export — the accessible HTML is the deliverable). Returns the distribution verdict, before/after scores, and every fidelity/honesty disclosure. Sends document content to the Gemini API.

Independent ISO 14289-1 (PDF/UA-1) validation of a local PDF using the packaged veraPDF CLI and local Java. It fails closed if that offline validator is unavailable; it never silently uses the CDN-backed browser validator. Run it on a -tagged.pdf produced by remediation or on any PDF. Needs NO Gemini key, paid Worker, institution account, AlloFlow service, or document upload; writes nothing. Typically 30-120s including JVM startup. This is a DIFFERENT artifact from the remediation score: the score judges the accessible-HTML content; this judges the exported PDF bytes.

Remove named strings (student names, emails, phone numbers, IDs) from an accessible HTML document, then VERIFY the removal: the pipeline re-scans its own output and reports `clean` plus any `leaks` that survived. Deterministic — needs NO Gemini key and nothing leaves the machine. Use before sharing a document that contains student data. A redaction that silently missed something is worse than none, which is why the verification pass is part of the result rather than an optional extra.

Report both operating modes: the tools available without a Gemini key/account/Worker and whether the optional full AI pipeline has its key, Playwright/Chromium, and modules. Call this first. Read-only; launches nothing, spends nothing.

Cancel a queued job (it never starts) or the running one (its browser context closes and in-flight AI calls die within seconds). Output files already written stay on disk.

The diagnostic snapshot for a run: the per-call ledger (outcomes, timings, byte counts, retries, models), throttle events, and the constants in force. Counts, timings, and enums only — never prompts, responses, or document text. Pass job_id for a background job; omit it for the most recent run in this server session. Use this to debug a slow, failing, or rate-limited run. Read-only.

The completed job's full summary (verdict, scores, honesty disclosures, output file paths; per-file summaries for batch jobs). Available only once remediation_job_status reports completed. Read-only.

Status of a remediation job (queued/running/completed/failed/cancelled) plus the most recent pipeline telemetry lines — throttle waits show up here, so a slow job is distinguishable from a stuck one. Read-only.

Run the REAL remediation pipeline end-to-end in headless Chromium against a scripted local model and a generated one-page PDF, then report which stage worked. Needs NO Gemini key and spends NO quota (nothing leaves the machine; the scripted model is a loopback server), writes no files you keep. Takes roughly 20-60s. Use this when remediation_capabilities says ready but real runs fail, after installing or updating the connector, or to tell a broken install apart from an API-key/quota problem: a failure here names the stage (assets / browser / module-boot / ownership-gate / audit-contract) and is never about your key.

Download the Chromium browser binary the pipeline runs in (a one-time ~200MB download via Playwright, 1-5 minutes). Call this when remediation_capabilities reports chromiumInstalled: false — typically right after installing the packaged connector. Idempotent: returns immediately if Chromium is already installed. Writes only to the Playwright browser cache; needs no Gemini key.

Prove the configured Gemini key WORKS rather than merely existing, by listing models — the cheapest authenticated call. Sends NO document content and spends no generation quota. Call this whenever remediation_capabilities reports a key but Gemini-powered tools fail, and after setting up a key for the first time. Distinguishes: no key configured, valid, valid-but-quota-exhausted, invalid (revoked/mistyped/wrong API), and unreachable (offline — key untested, not proven bad). When no key or an invalid key is found it returns exact setup instructions. Never returns or logs the key value.

Produce a plain-language version of an accessible HTML document, preserving its structure and reading order. This is the UDL move that most changes who can use a document: an accessible PDF that is written at a graduate reading level is still inaccessible to many readers. REQUIRES a Gemini key. The simplified text is model-generated and should be reviewed by someone who knows the subject before it replaces the original.

Transcribe a local audio or video file into an accessible transcript payload. Modes: 'speech' (spoken words), 'visual' (what is shown on screen), 'dual' (both, separately), 'synthesis' (a combined narrative). Long recordings are segmented automatically by the pipeline's large-file machinery. Supports .mp3 .m4a .wav .aac .ogg .flac .mp4 .mov .webm .mpeg. REQUIRES a Gemini key and sends the media to the API. This is how a recorded lesson becomes a document the rest of these tools can make accessible.

Translate an accessible HTML document into a target language while preserving its structure: headings, lists, tables, and reading order survive, and embedded data-URI images are swapped for placeholders during translation so they are never sent to the model and come back intact. Long documents are chunked on tag boundaries. REQUIRES a Gemini key. Use to produce the same accessible document for multilingual families or students.