Release notes¶

Unreleased — Clickable dashboard errors¶

• Each issue in the New Run error summary links to its field: selecting it opens the wizard step that contains the field, scrolls to it, and focuses it. Fields named by a server error are outlined in red, and the wizard opens on the step of the first listed issue.
• Subscription model-availability errors now name the Stage 1 or Stage 2 model field instead of a generic Model entry with no matching control.
• Subscription model menus list the provider's default and recommended models first (Recommended by Google), so Google accounts now preselect Gemini 3.8 Flash instead of Gemini 2.5 Pro, which Google lists but refuses with HTTP 503.
• Before the Review page, each chosen subscription model gets one tiny test request; a model the provider refuses is rejected on its field instead of failing the run after repeated retries. A blank stage model takes the first catalog model, starting with the provider's default, that answers.

Unreleased — Stage 1 typography on by default¶

• Inference now asks Stage 1 to wrap confidently bold text in <b>…</b> and italic text in <i>…</i> by default, in the CLI, YAML, and dashboard.
• Disable it with --no-stage1-typography, pipeline.stage1_typography: false, or Plain text under Stage 1 typography in the dashboard's Pipeline step.

Unreleased — Dashboard upload recovery¶

• Browser uploads remain in a local server-side draft across validation errors and same-tab reloads. The dashboard shows retained filenames and supports replacing or removing draft uploads; inactive drafts expire after one hour.

Unreleased — Partial Dictionary Profile context¶

• Dictionary Profile questions are individually optional.
• Supplying a headword or target language requires its matching script.
• Page layout, information types, and character inventory can be supplied independently; leaving the rest blank no longer blocks a run.

0.2.0 — Dashboard and extraction redesign¶

Released October 24, 2026.

MUDIDI 0.2.0 brings the redesigned dashboard, parallel runs, page-local MDF conversion, and an Evaluator–Editor loop. Subscription response handling and dependency security fixes are included.

Page-local Stage 2¶

• Stage 2 now parses each page locally. A page's MDF holds exactly the lines printed on that page; an entry that crosses a page break is split between the two page files instead of being completed from the next page.
• New dictionary.mdf.txt joins the page files in order, and dictionary.mdf.report.json lists the pages that start mid-entry.
• Stage 2 Pass 2 sends one page image instead of three. The previous page is supplied as the last 1,500 characters of its transcript; the next page is no longer sent.
• Pages run eight at a time by default (--batch-size 8).
• Oversized page images are compressed once and reused across stages.

Agentic loop: Evaluator and Editor¶

• The loop is now an evaluator-optimizer. The Evaluator chooses an explicit action (accept, targeted_edits, full_redo, or reject) and lists each edit as a line, the exact current text, and the replacement.
• The Editor (previously "rewriter") verifies every proposed edit against the source and approves or refuses it with a reason. Approved edits are applied as exact replacements; the Editor no longer re-writes the page.
• The Editor's verdicts are fed back to the Evaluator on the next round.
• The loop stops early when it detects no progress: a refused edit proposed again, two rounds that change nothing, or a page returning to an earlier version.
• Maximum Iterations replaces "Maximum correction iterations" and defaults to 3.
• Deterministic patches and Concrete retry evidence are removed from the dashboard. The matching YAML keys and CLI flags are still accepted but ignored.
• Loop records use new fields (action, edits, redo_reason, per-round Editor verdicts) and stop reasons (invalid_decision, no_progress, oscillation).
• Subscription logins wait and retry when the provider reports that it is overloaded or rate limiting, instead of failing the page.
• dictionary.mdf.txt leaves out a page whose file holds no MDF (for example a model refusal) and lists it under pages_without_mdf.

Parallel runs¶

• The dashboard no longer limits you to one working run. Start as many as you like.
• Active run is now Active runs: a list of every run in progress with its stage and progress, including runs waiting for parsing-guide review.
• When a worker process crashes, the run page now shows the worker's last error line instead of a bare "failed" status.
• A run is refused only when another in-progress run writes to the same output folder. The Review page names that run.
• The Review page and the run page warn when another working run uses the same provider login, because those runs share one rate limit.

Dashboard redesign¶

The local dashboard has a new look and a reorganized New run flow. Routes, form fields, saved presets, and run data are unchanged.

• New run is now a six-step wizard: Input, Pipeline, Context, Model, Agentic, and Review. The new Context step holds the Dictionary Profile, MDF parsing-guide inputs, and an Advanced: stage instructions toggle.
• The interface uses the Modernist Ochre theme with light and dark modes and self-hosted fonts.
• Every run page shows the same workspace tab bar; views that are not available yet are greyed out instead of hidden.
• Native installs add Choose folder… beside the output directory, which opens the operating system's folder dialog.
• Agentic verification is renamed Agentic loop, with an explanation of what it does and what it costs.
• Dashboard runs always add the bundled SIL MDF manual text to Stage 2 Pass 1; the manual menu is removed.
• Request and upload size limits are off by default and opt-in through --max-request-bytes and --max-upload-bytes.

Reliability and maintenance¶

• Subscription adapters support PDF attachments and retry malformed structured replies. Gemini guide discovery and agentic checks accept the provider's looser JSON responses.
• OpenAI subscription discovery lists every model the account can use.
• Worker crashes expose the last error line, and runs waiting for guide review remain visible under Active runs.
• Evaluation results and tooling are migrated, with Python replacing the previous R table generators.
• Locked dependencies are updated to address published security advisories.

Upgrade notes¶

• Per-page MDF no longer completes an entry using the next page. Use dictionary.mdf.txt for the joined dictionary and dictionary.mdf.report.json to inspect boundaries and pages without MDF.
• Consumers of agentic JSON records must handle the new action, edit, Editor verdict, and stop-reason fields described above.
• The retired deterministic-patch and concrete-retry settings remain accepted but have no effect.
• Leaving parsing-guide sample pages blank now uses every selected dictionary page. For large dictionaries, choose representative pages explicitly to control the size and cost of the discovery request.

0.1.1 — Google subscription workflow¶

MUDIDI 0.1.1 makes Google subscription login and model selection work entirely from the local dashboard.

Google authentication and models¶

• Log in and out through Google Antigravity's browser OAuth without requiring an Antigravity CLI login.
• Store the resulting access and refresh tokens only in MUDIDI's encrypted local subscription store.
• Refresh live subscription model catalogs on demand and invalidate cached results when login or logout changes the active account.
• Put Gemini Pro models before Flash models, preserving newest-first provider order within each family.
• Show the account as Google subscription in the dashboard.

Compatibility¶

• Migrate legacy saved presets containing models.temperature before strict validation.
• Preserve API-key billing as an independent authentication mode.

0.1.0 — Subscription inference¶

MUDIDI can now run production dictionary digitization through authenticated OpenAI, Google, or Claude subscriptions. API-key inference remains available as a separate billing mode.

Subscription accounts¶

• Sign in and sign out from the local dashboard or mudidi auth.
• Keep provider credentials in a dedicated encrypted local store.
• Use independently authenticated providers without exposing access tokens in configuration, logs, subprocess arguments, or browser storage.
• Refresh expired sessions once and fail with actionable authentication guidance when reauthentication is required.
• Make Subscription billing the first and default billing choice for new dashboard runs; API-key billing remains available below it.

Live model selection¶

• Load stage-compatible models from the authenticated account rather than a fixed OpenAI, Gemini, or Claude list.
• Sort Google subscription models with Pro before Flash while preserving newest-first order within each family and for other providers.
• Show only the reasoning levels supported by the selected model, including adaptive Claude and effort-qualified Gemini models.
• Configure Stage 1, Stage 2 Pass 1, Stage 2 Pass 2, evaluator, and rewriter models independently where the workflow permits it.
• Preserve entitlement-only split-model selections in saved presets while the live catalog refreshes.

Production workflow¶

• Run subscription-backed inference from both the local web dashboard and mudidi run.
• Review and approve the generated MDF parsing guide before Stage 2 conversion.
• Resume compatible CLI output safely; changed instruction attachments require an explicit overwrite.
• Cancel queued or active web runs without leaving page output or usage attributed to work that never completed.
• Pass --alphabet PATH to enable the supplied character inventory automatically.
• Migrate older saved dashboard presets that contain retired configuration fields instead of failing during dashboard startup.

What changed¶

• Dashboard preset storage now runs schema migration 5 for existing local databases.
• Legacy models.temperature values are removed before strict configuration validation, so old presets no longer crash dashboard startup.
• The migration preserves current preset settings and keeps new YAML and API configurations strict; temperature remains rejected for newly submitted configurations.

Provider adapters¶

• OpenAI: Codex account authentication, account-scoped model discovery, structured output, image input, reasoning controls, and usage normalization.
• Google: direct Google Antigravity OAuth and Cloud Code routing, canonical Gemini model aliases, effort routing, failed-run retry, and project-aware requests. The official Antigravity installed-app registration is bundled; the MUDIDI_GOOGLE_OAUTH_CLIENT_ID and MUDIDI_GOOGLE_OAUTH_CLIENT_SECRET variables are optional overrides for compatible Antigravity registrations.
• Claude: OAuth-backed Messages requests, paginated model discovery, image input, structured output, adaptive or extended thinking, and typed failure handling.

Verified release boundary¶

The release was smoke-tested with page 34 of the Carolinian-English Dictionary through all three authenticated providers and both supported surfaces. Coverage included complete Gemini extraction, OpenAI and Claude transcription, CLI split-model execution, agentic verification, cancellation, guide approval, editing, artifacts, history, usage, presets, and resume behavior.

Three defects found during that smoke were repaired before release:

1. saved subscription presets losing entitlement-only split models during catalog startup;
2. complete overwrite runs issuing a duplicate Stage 1 request for the Pass 1 sample page;
3. explicit CLI alphabet files remaining disabled in the resolved runtime.

The release candidate passed 1,383 repository tests with 10 intentional skips, the strict MkDocs build, generated-reference verification, Ruff, and live post-repair web and CLI checks. The subsequent legacy-preset migration regression brought that historical suite to 1,385 tests.