Every pass was English-shaped: CheckpointMessages took the text and the tone and
nothing else, so there was never a language decision to get wrong. On the live
build two pt-PT sentences drew no cards at all — Petal read the Portuguese, said
nothing about it, and filed a mechanics note about the one English line.
The rule is two decisions reading different state. What gets corrected follows
the document. What language the explanation is written in follows the writer —
the half of her pair she is not learning, from users.direction — because an
explanation is teaching, and teaching lands in the language she reads most
easily. Those coincide for every account that exists today (learnerPairs is
{"zh"}), which is a fact about the roster and not about the design, so Target
keeps them apart. It carries a third language too: the collocation gloss is
addressed to her rather than to the document, and folding it into Explain would
have quietly moved it into English on every English document.
The document verdict is a proportion, not a presence — one Portuguese quotation
must not flip an English essay. Per sentence, three-way: pair, English, or no
answer. The third value is the load-bearing one; counting the undecided as
English is exactly what would hold a journal of short Portuguese sentences in
English forever, so the Latin pairs needed an englishMarkers list curated against
pt/fr/es as carefully as latinMarkers was curated against English. Hysteresis at
70/40 because a bilingual paragraph would otherwise alternate its cards' language
every few keystrokes, and hysteresis needs a yesterday — hence the column. Plus a
corroboration floor: a ratio computed over "Não. Eu." is 100% of nothing, and a
flip rewrites every card in the document.
The verdict folds into the chunk salt beside the tone, so a document that changes
language re-opens every sentence rather than serving back cards in a language it
no longer speaks.
checkpointSystemPrompt could not simply take a language — it opens by naming the
reader an ESL learner, and appending "explain in Portuguese" hands the model two
contradictory framings. Separate constants, sharing the JSON contract below the
framing. Both carry a "never translate it into English" line, which is the
instruction the model will most want to disobey. The English prompts are
untouched byte for byte, and a golden says so out loud.
Collocation deliberately did not move: its prompt is per-language knowledge, not
framing, and "natives usually say" for Portuguese is a claim Petal cannot back.
Not deployed and not smoked against a real model. The tests drive the real router
and a real DB; what none of them prove is how Qwen behaves on a Portuguese
document, in particular whether the never-translate line holds.
Claude-Session: https://claude.ai/code/session_01GJHNvirh7Hzhc9RL3HAvz7
Every `中文 · English` string moves out of ~29 components into
web/src/i18n: one Pack type, a verbatim zh pack, and two ways to read
it — usePack() for components, pack() for the modules that build a line
when something happens rather than when something renders.
Anything with a value in it is a function on the pack rather than a
template at the call site, English pluralisation included: word order
isn't universal, and a pack author has to be able to move the number.
The roster constants (tones, rewrite styles, export formats, companions)
keep only value + emoji, so a label can't drift from its key.
On the server, internal/llm/lang.go replaces "Simplified Chinese" in the
three prompts that actually name her language. pt-PT is spelled
"European Portuguese (pt-PT, never Brazilian Portuguese)" in the prompt
itself, and each Lang carries her word for "why" so the tutor prompt
still recognises the question when she asks it her way.
pair_lang reaches the model through the row-scoped query each handler
already ran — the one that proves she owns the document — rather than a
second lookup that could disagree with it.
Also records Phase 18's deploy: migration 0011 rehearsed against a copy
of the live VPS database, then applied for real.
The checkpoint model sometimes "flags" a correct sentence and echoes it
verbatim as the replacement, producing a card whose before/after are
identical and whose explanation says it's already fine — a suggestion
that suggests nothing. ParseCheckpoint already dropped empty-original
items; also drop these no-ops. Awareness-only families (voice) carry an
empty replacement, so the guard only fires on the edit families.
Claude-Session: https://claude.ai/code/session_016Yr6jELuRc7hyzYLccQKZd
The grammar checkpoint capped num_predict at 1024, but qwen3.5:9b ignores
the prompt's "up to 5 issues" and emits ~17-20 suggestions (~2000 tokens)
on a 300+ word doc. The output hit the cap mid-array (done_reason=length),
the JSON never closed, and ParseCheckpoint found no parseable object -> a
502 in ~20s on every long doc (1024 tok @ ~50 tok/s, not a network
timeout). The repeated failures made the writing-assist helper look
permanently asleep.
Fix:
- Raise the cap to checkpointMaxTokens=4096. It is a ceiling, not a
target -- the model stops at its JSON close, so shorter docs are
unaffected; only genuinely long outputs use the headroom.
- Make ParseCheckpoint salvage the completed {...} suggestion objects
from a truncated array (refactor extractJSONObject onto a shared
firstBalancedObject scanner), so an over-long doc degrades to partial
feedback instead of a hard 502.
Verified live on millenia: 'The Missing Key' (382 words) now returns 200
with 17 suggestions in ~38s, previously 502 every time.
Claude-Session: https://claude.ai/code/session_016Yr6jELuRc7hyzYLccQKZd
A paste fires exactly one grammar checkpoint, and a failed one never retried
until the next keystroke — stranding the writer on "Petal's helper is resting"
after a paste. Long docs make it worse: their 15-25s checks have a wide window
to catch a transient 502 from the shared Ollama (co-tenant apps load other
models and evict the 9B). A failed pass also burned the per-document rate-limit
slot, so a retry within 30s hit the throttle path and got an empty set back.
- llm.RateLimiter.Release rolls back a slot when its pass fails; Allow now
returns the recorded timestamp so Release only frees its own slot.
- suggestions.runPass releases the slot on LLM failure before returning 502.
- useCheckpoint auto-retries a failed checkpoint with backoff (3/12/35s),
keeping the breathing dot up and only flagging "resting" once retries exhaust.
Claude-Session: https://claude.ai/code/session_016Yr6jELuRc7hyzYLccQKZd
Four enhancements to make the editor fit real school usage:
- Per-document tone (academic/professional/casual/humorous/creative/
persuasive/general): new documents.tone column (migration 0002), threaded
through the docs API, a bilingual ToneSelect dropdown on the title row, and
injected into the grammar-checkpoint LLM prompt so advice fits the register.
The voice pass stays tone-agnostic.
- Right-click word lookup: a new offline `lexicon` package serves definitions
(Wordset, modern ESL-friendly glosses) and synonyms (WordNet synsets first,
then frequency+stopword-ranked Moby for breadth) from gzipped embedded data,
behind /api/word/{word} with light morphology. The WordCard popover shows the
definition and tappable synonym pills that swap the word in place.
- Expanded writing stats: clicking the word count opens a StatsPanel with page
count, sentences, paragraphs, reading time, average word length, word variety,
and Flesch-Kincaid reading level — all computed client-side.
Claude-Session: https://claude.ai/code/session_016Yr6jELuRc7hyzYLccQKZd
Backend (internal/llm): backend-agnostic LLMClient interface + factory
with vLLM (OpenAI-compat) and Ollama (native) clients, each Complete +
Stream. prompts.go holds the checkpoint and Ask Petal templates;
checkpoint.go salvages JSON from model output (brace-matched), enforces a
per-doc 30s RateLimiter, and truncates the doc to a latency cap.
internal/suggestions: POST /api/docs/:id/check runs a checkpoint and
replaces the doc's pending suggestions in one tx (accepted/rejected kept
as history); GET /api/docs/:id/suggestions lists pending;
POST /api/suggestions/:id/{accept,dismiss} resolves one. Throttled checks
return the current set rather than erroring.
Frontend: useCheckpoint (4s debounce, loads existing on open, stale-guard
tokens); SuggestionHighlight renders ProseMirror decorations re-anchored
by the `original` string on every doc change (not stored marks), with
precise textblock-offset→PM-position mapping; SuggestionCard shows the
type tag + diff + explanation and applies the replacement in-editor on
accept; breathing rose checkpoint dot in the StatusBar; fade-float +
breathe animations.
Tests: llm parse/rate-limit/truncate; suggestions full flow + rate-limit
over httptest with a stub client. Smoke-tested end-to-end against a fake
vLLM endpoint (anchoring verified) and the LLM-unreachable 502 path.
Claude-Session: https://claude.ai/code/session_016Yr6jELuRc7hyzYLccQKZd
prosolis