Text-to-speech, MCP capability guides, and a two-tree workflow catalog - #41

- Removed the MCP proposal document as it is no longer needed.
- Added a new task for speech generation using Hugging Face transformers, allowing the generation of audio from text with a specified voice preset.
- Introduced AudioTrack class to encapsulate generated audio and its sample rate.
- Updated media_arguments to handle AudioTrack and AudioVideo types.
- Created workflows for generating speech and integrating it into video generation, ensuring that the audio sample rate is preserved.
- Added tests for the speech generation functionality to ensure correct behavior and integration.

Separates the catalog into templates (teach a pattern) and model configs
(record a hardware fact), which have been sharing one flat list of 121
entries. Measured: 54 of 84 non-archive workflows share a feature
signature with another, and the clusters are checkpoint iterations rather
than feature variations.

Records the compatibility constraints a move runs into - the step cache
survives on id, run-directory identity does not survive a path change -
and defers the derived shape index until after the restructure.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Backward compatibility is not maintained: existing outputs may be
orphaned and references into them may break, and no alias or migration
shim is carried for a one-time move.

What replaces it is stricter - when this lands, every reference in the
repo resolves. Enumerates the reference classes with counts and what
enforces each, which surfaced one gap: 18 distinct prompt: references
across 20 workflows that nothing checks, and that collapsing will rewrite.
Closing that gap is now part of the work.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Seven tasks: close the prompt: reference gap first as a safety net,
inventory all 121 workflows and stop for approval, then build
templates/ and models/, prune, repoint the docs, and prove every
reference class resolves.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Sub-workflow paths, type names and constants each have a test that they
resolve; prompt: references had none, and they fail on the GPU rather
than in CI. Added ahead of the catalog restructure, which rewrites the
20 workflows carrying them.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The first draft argued from a crude feature signature and overstated the
redundancy. Reading found it concentrated in archive/, which is a pure
model gallery, while minimax/ and ltx2/ are a curriculum where each file
demonstrates a conditioning mode or mechanism nothing else does - those
stay. gyre/ is a test project's production history; only the mechanisms
that generalize earn templates.

Also records two ids used by two workflows each, which collide in the
step cache today and are resolved by emptying archive/.

Expected scale is now 45-50 deletions rather than a halving.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

A verdict per file - template, model config, collapse target or delete -
with the id each surviving file must keep. Awaiting approval; nothing is
moved or deleted until it has it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

One approved exception to the keep-ids-verbatim constraint: the
text-to-image template takes the id 'text-to-image' rather than the
'test_job' it would inherit from sd15.json. Costs one workflow's
step-cache entries.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

One file per capability, with six checkpoint variations collapsed into
argument sets in each template's description where the pipeline class
matched. Every moved file keeps its id, so the step cache survives.

Two deliberate id deviations, both narrow. The text-to-image template
takes 'text-to-image' rather than sd15.json's 'test_job' (approved).
The three templates replacing GYRE's editing passes are authored fresh
rather than moved and take new ids: the originals referenced
workflows/gyre/assets/, which was never committed, so they could not run
as written and have no cache entries to preserve. The rest of the GYRE
project goes with them - it is a test project's production history, not
a set of examples.

Doc links are updated in a later commit; test_docs_links fails until then.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Each names the template it configures, so a tuned 24GB config is findable
as an instance of a pattern rather than as another undifferentiated
catalog entry. All nine configure templates/text-to-image: the collapse
rule stops at a pipeline class boundary, and FluxPipeline, Flux2Pipeline,
Krea2Pipeline and ZImagePipeline are four of them.

The ltx2/ and minimax/ READMEs move with the families they document.
Ids preserved.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Every removed file was a further checkpoint through a pattern the tree
already demonstrates. Git history is the archive, so archive/ stops being
a graveyard that costs listing space. 121 workflows become 73.

compose-workflows.json named two archive workflows as its defaults, which
would have broken silently - it now points at surviving files.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

All 66 broken workflow links, plus 75 whose text still named the old
filename, plus the REPL examples that loaded flux/FluxDev.

Fourteen links whose targets were deleted needed judgement rather than a
mapping. In RECIPES_24GB the per-model VRAM prose is knowledge no template
carries, so it stays and only the dead example links go, replaced by a
pointer to the template to author from. The IP-Adapter, LoRA and
quantization examples move to the surviving template or model config.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

templates/ teaches a pattern, models/ records what makes a checkpoint fit
and names the template it configures. The workflows guide leads with the
layout, and the MCP guide summary says so - that is the first thing an
agent matching a request's shape needs.

Also corrects the plan's own id check, which excluded only DELETE rows and
so would fail on the six COLLAPSE rows whose files were folded away.

Every reference class resolves: markdown links, sub-workflow paths, type
and constant references, and prompt: references.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The restructure regressed browsing without touching the UI. FolderGroups
grouped by the first path segment only, so the new layout collapsed to two
groups - templates (64) and models (9) - hiding the ltx2 and minimax
families entirely, where the old layout showed six. Gallery entries
regressed the same way: their names are <workflow identity>/<file> and the
identity is now itself a path, so every run grouped under 'templates'.

Grouping by everything but the last segment restores the structure and
fixes both pages, so it lives in one shared helper rather than being
recomputed per page.

The listing also carries 'configures' now, and a model config's card says
which template it is a tuned instance of - without it the two trees look
identical in the one place a person actually browses them.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The follow-on the restructure deferred, written up as a proposal rather
than left in the spec's 'what this unblocks' paragraph.

Derive a small shape vocabulary in workflow_details, which already parses
and caches every workflow, with an optional declared override where the
derivation is wrong. Report observed runtime from the jobs table.

Records the join problem as confirmed rather than assumed: jobs.workflow
stores the workflow's id, not its catalog path, and ids are not unique by
construction - two shared 'sd35' until archive/ was emptied.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The field was added to WorkflowsPage's local copy of the detail shape but
not to the api.ts type it is assigned from, so npm run check failed on the
incompatible index signature. Optional there, matching prompt_refs: the UI
can be talking to a server older than this field.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The Playwright suite drives the real catalog by name, so the renames broke
five specs: ZImage, flux/FluxDev, tasks/ImageToText and FluxFill no longer
exist. Routes now name the new paths, headings expect the full catalog
name and breadcrumbs the leaf.

The filter assertion is anchored as /^inpaint / rather than /inpaint/,
which matches both the inpaint template and segment-and-inpaint.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…confined

An unconfined CLI run (workflow_from_file with no workflow_dir) had only the
'..' regex standing between a relative sub-workflow path and the filesystem
above it. Normalising the path before validation (previous commit) collapsed
that regex's usefulness away, so a path like '../../../../etc/passwd.json'
would resolve and pass. Confine such a path to the nearest ancestor directory
literally named 'workflows', reusing the approach dw.runs.workflow_identity
uses for output naming (dw.workflow.catalog_root_dir), falling back to the
referencing file's own directory when there is no such ancestor.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…httpx

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Audio-only pipelines now record their vocoder or VAE rate the way LTX-2
already did, and get_artifact_list wraps each item so a workflow need not
declare one. Follow-up: a modular pipeline's dict output (music.json) still
carries no rate of its own.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…o is saved as audio

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

… drops

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…osts