A run's detail only carried the last dozen items, so a model that failed early
scrolled out of reach: run 21 reported "1 failed" with no way to learn which
model or why. The detail now carries every failed item, listed in full.
"Retry failed" existed but could never appear: it needs a run that is finished
with failures, and such a run is 'done', which neither canResume nor resume()
accepted. Both now do; resume() already no-ops when nothing is pending, so a
clean run is unaffected.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The Red panda kept failing with "The AI reply wasn't the JSON asked for", and
nothing showed what came back. Now:
- the prompt includes an example object with just the requested fields
(DeepSeek's JSON mode can otherwise send back an empty reply)
- JSON is found inside code fences, notes, or braces after it
- a bad reply is asked again up to twice: with more room when it was cut off,
and the last time without JSON mode; max_tokens starts at 1500
- the error says whether the reply was empty or cut off and quotes its start
- tokens spent on failed attempts count towards the run's cost
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- Runs can now do any mix of: tags & descriptions, clearer names, and
collection suggestions. A run without tags doesn't mark models as tagged.
- Collection ideas reuse your collections and earlier ideas where they fit and
wait on the AI page, grouped by name, until you create/add or dismiss them.
Dismissed ideas aren't suggested for that model again; spelling variants
group together under the capitalised spelling.
- A reply that isn't the JSON asked for is asked again once; both requests
count towards the run's tokens and cost.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A run option asks the AI for a clearer name only when the current one is messy
(file-name leftovers, typos, site boilerplate). The previous name is kept, the
run feed shows what was renamed, and the model page offers Undo. Giving a model
a name yourself clears the undo.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The model page's tag list wasn't a wrapping row, so the last tag could break in
half, and four action buttons crushed onto two lines each. AI runs also never
woke the background worker, so new AI tags and descriptions weren't searchable
until something else changed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- AI tagging through any OpenAI-compatible API: DeepSeek by default, or
OpenRouter, OpenAI, or a local server (Ollama, LM Studio). Settings → AI
with presets, prices and "Save & test connection"; the key never leaves
the server except to the provider.
- Runs over untagged models, all models, or one model from its page: live
progress, tokens and cost, stop/resume (retrying failures), an optional
spending limit, and paused rather than resumed after a restart. Adapts to
models without JSON mode or image input; stops on a rejected key or no
credit. Tags are only added; descriptions filled only where empty unless
asked.
- Search by meaning: local bge-small-en-v1.5 embeddings (transformers.js),
indexed in the background whenever a model's text changes; Words/Meaning
switch on the search box. Stored as REAL[] in Postgres, no pgvector needed.
- Automatic tags: strict whole-word rules (part types, mechanisms, materials,
techniques) over names, file names and archive contents, plus your own
word -> tag rules; on/off (optionally removing its tags) and re-apply.
- Tags remember where they came from (user / rule / ai); edits keep that,
rule and AI passes never touch your own tags, and the model page marks them.
- Docker: ships the embedding model and only Linux onnxruntime binaries.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
