Reported from the phone: "I have my main language set as Portuguese and I say
I'm learning English but yet Petal presents the Ask Petal advice in both
sections as Portuguese."
Nothing was wrong with targetFor again. A Portuguese document by a Portuguese
writer is explained in Portuguese, which is the whole point of Phase 28. The
card was right. What was wrong was the tap underneath it: /suggestions/{id}
/translate answered "" for exactly that case, on the reasoning that an
unasked-for English rendering of an explanation she can already read is not a
seed but noise. That reasoning had the writer facing the wrong way. She is
learning English. The half she is *practising* is the half worth a tap, and the
bubble sits directly beneath the explanation inside the same card, so answering
"" left her with Portuguese, the same Portuguese again, and no English anywhere
on the card. targetFor's own comment promised the other language stays one tap
away in both directions; only one direction had ever been built.
So the endpoint keeps the one rule it always claimed: render into whichever
half the explanation is not already in. English explanation into her language,
as before; her language into the English she is learning, which is new. Both
ends of that are now parameters — TranslateMessages took the source language
for granted as English because until Phase 28 it always was. The zh prompt is
unchanged byte for byte, which its test still pins.
The client fallback was the same symptom from a different cause and would have
survived the server fix: an empty answer, or an unreachable model, seeded the
bubble with the explanation itself — a verbatim repeat of the line two above
it, which reads as Petal replying in the language the tap was pressed to
escape. With the endpoint always having somewhere to go, empty now means only
that the model didn't answer, so the panel opens with no bubble at all and the
input where she can ask. AskPetal no longer takes the explanation as a prop; it
never needed anything but the id.
The test that pinned the refusal now pins the rendering, and carries the report.
Claude-Session: https://claude.ai/code/session_01KGACAtTPjvZ2PipDZ5qD99
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
The tutor prompt said "never mix languages in a single response" and
mirrored the language of the question, so asking in English — which she
does, because she is practising — returned the one explanation surface
that gives nothing in her own language. It now answers in both, pair
language first, halves separated by a blank line.
Which half is the safety net and which is the lesson depends on who is
writing: the pair is (English + X) and Petal is used from both ends, so
the prompt asks for both and says it doesn't know which way round.
The split is a rendering nicety, never a parse the reply depends on: a
half-streamed reply is all one half, a model that ignores the
instruction renders as one block, and nothing is ever dropped.
For the height, the first attempt clamped the box to the room left below
the anchored card so it could never overhang — measured, that gave 176px
against a 442px answer, worse than the 220px it replaced. The card's own
chrome spends ~290px of an 810px window, so "fits below the word" and
"room to read" are not both available. The ceiling is now a flat 50vh and
the overhang is made navigable instead, per item 4: the card reports its
reach like the rail already does, the column grows, and the page can
scroll to the actions below it.
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.
Qwen3-family models reason by default and prepend a plain-text preamble
ahead of the answer — not a <think> block, so it cannot be stripped
after the fact. Every Petal pass parses a JSON object out of the
completion, so an unsuppressed preamble fails the parse outright.
Send chat_template_kwargs {"enable_thinking": false} on every request,
matching the unconditional think:false the Ollama backend already sends.
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
Phase 12 — collocation coach: a third suggestion family for gentle
"natives usually say…" hints on non-native word pairings, reusing the
existing runPass/pendingScope/rail machinery.
- llm/collocation.go (RunCollocation, 25s floor, reuses ParseCheckpoint)
+ collocationSystemPrompt/CollocationMessages (warm, Mandarin gloss,
defers grammar/spelling to the grammar family)
- migration 0005 rebuilds the suggestions table to extend the type CHECK
(SQLite can't ALTER a CHECK)
- collocationScope + CollocationLimit + POST /{id}/collocation
- fix: grammarScope was `type != 'voice'` and would wipe the new
collocation flags; now `type NOT IN ('voice','collocation')`
- frontend: --color-blossom, "Make it sound natural 🌸" pill,
collocating/runCollocation in useCheckpoint, StatusBar dot
Phase 13 — vocabulary garden: capture looked-up words and surface them
for gentle spaced repetition.
- new internal/vocab package: migration 0006 (vocab_words, SM-2-lite
columns, doc_id ON DELETE SET NULL, UNIQUE(user_id,word)),
scheduler.go (Leitner ladder 1/3/7/16/35 then geometric; gentle
"again", no streak-shaming), handlers (capture-upsert/list/due/
review/delete, owner-scoped, SQLite-side datetime math)
- auto-capture on word lookup (dictionary-known words only, captures
the surrounding sentence + doc_id) + 🤍/💚 toggle on WordCard
- GardenPanel: blossom grid (bloom by reps), flashcard review (sentence
blanked, flip, again/good/easy, recognition↔production), sleepy-kitten
footer; opened from a global 🌷 header button
Tests: TestCollocationPassCoexists, vocab scheduler + handlers, db CHECK
extended. go build/vet/test + tsc + vite + vitest (51/51) clean;
migration verified against a copy of the live DB; live backend smoke
walked the full vocab lifecycle + the warm-502 collocation path.
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
Surface every outstanding suggestion as a card in the right-hand
whitespace, vertically aligned to the text it flags — so the writer sees
the whole queue at once instead of hovering each highlight. Cards stack
with collision avoidance, link both ways with their highlight (hover/click
↔ soft text wash, driven through the decoration plugin so it survives
edit repaints), and carry the same Accept / Dismiss / Ask Petal actions.
The rail is a progressive enhancement: it mounts only when there's room
beside the editor, otherwise the existing inline hover card is unchanged.
Stacked cards that reach the bottom-right corner tuck behind the
companion mascot (z-order).
When a card is expanded, the Ask Petal bubble now opens with the
Simplified-Chinese translation of the explanation (the English stays in
the card body) instead of repeating the same text twice — a new
POST /api/suggestions/{id}/translate one-shot LLM endpoint, loaded
lazily on open with an English fallback.
Verified live against the local LLM via the uitest harness: rail
stacking, hover↔text wash, expand/Ask Petal, accept-from-rail, narrow
fallback, and the Mandarin bubble.
Claude-Session: https://claude.ai/code/session_016Yr6jELuRc7hyzYLccQKZd
Inline Chinese gloss (offline) and a "say it more naturally" / tone-rewrite,
the two ESL features for the Mandarin-speaking writer.
Gloss: embedded English→Chinese dictionary (gloss.json.gz, 57k common words
built from ECDICT via scripts/build_gloss.py). lexicon gains Gloss()/Result.Gloss
and a lightweight GET /api/gloss/{word}; the right-click WordCard leads with the
中文; GlossTip shows it on a 350ms hover (reuses wordAt, so CJK is never glossed).
Offline + instant, works with the LLM down.
Rewrite: selecting text pops a SelectionBubble (✨更自然 + the tone vocabulary);
picking a style calls POST /api/docs/:id/rewrite (llm.RunRewrite, stateless,
owner-scoped) and shows a RewritePreview (original→rewrite, accept/cancel/retry).
Accept applies it in-editor.
Tests added in lexicon and suggestions. go build/vet/test, tsc, vite all clean;
live smoke vs a fake vLLM verified gloss + rewrite + 400/404/502 paths.
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
Tier-1 voice-consistency pass: whole-document LLM review surfacing passages
that read tonally out of place (formal/over-polished/paraphrased-too-closely),
as honey-decorated `voice` flags with no correction (awareness-only).
- internal/llm/voice.go: RunVoice sends the whole document (no TruncateDoc),
MaxTokens 2048, 20s per-doc floor (VoiceInterval). Standalone voice prompt
in prompts.go (not bundled with the grammar checkpoint, per spec).
- internal/suggestions: POST /api/docs/:id/voice. replacePending is now
family-scoped (pendingScope) so grammar and voice never clobber each other's
pending flags; both passes return the unified pending set. check/voice share
one runPass helper. TestVoicePassCoexists covers both directions.
- Frontend: api.voiceDoc, useCheckpoint voicing/runVoice, honey "Check my
voice" toolbar pill, breathing honey dot in StatusBar.
Claude-Session: https://claude.ai/code/session_016Yr6jELuRc7hyzYLccQKZd
Conversational follow-up on a suggestion, streamed token-by-token.
Backend (interface-only; handlers never touch a concrete LLM client):
- internal/llm/chat.go: StreamAskPetal with conversational sampling
(max_tokens 512, temp 0.7, rep 1.15, top_p 0.92, stop "\n\n\n"),
reusing AskPetalSystemPrompt + TrimHistory.
- internal/suggestions/chat.go: POST /api/suggestions/:id/chat. One
user-scoped join loads the suggestion + parent content_text;
surroundingParagraph extracts the \n\n-bounded paragraph at from_pos
(whole-doc fallback when unlocated) and injects it server-side.
Streams event: token / event: done SSE frames with JSON-encoded data
so token newlines can't break framing; real http.Flusher per chunk.
LLM-unreachable -> 502 before SSE headers; unknown suggestion -> 404.
Frontend:
- streamSuggestionChat: fetch + ReadableStream SSE parser (not
EventSource, needs POST), abortable.
- AskPetal.tsx: whole conversation in component state (no persistence,
cleared on close), Petal's first bubble pre-seeded with the
explanation, rose/lavender bubbles, CJK font stack on the bubbles
only (Note #17), streaming caret.
- SuggestionCard "Ask Petal" pill pins the card open while chatting
(hover-close suppressed, click-away closes) and widens it to 340px.
Tests: chat_test.go covers streamed-text concat + done event,
server-side context injection on the system message, sampling params,
404, and surroundingParagraph. go build/vet/test clean, tsc clean,
vite build OK. Live SSE smoke-tested against a fake streaming vLLM:
tokens flushed individually through the chi middleware stack, done
terminator, 502 on LLM-down, 404 on unknown suggestion.
Claude-Session: https://claude.ai/code/session_016Yr6jELuRc7hyzYLccQKZd
Qwen 3.5 — the spec's recommended model — is a reasoning model. With
thinking on, Ollama streams its chain-of-thought into a separate
`thinking` field and hits num_predict before emitting any answer into
`content`, so Complete() got an empty string and the checkpoint failed
with "no JSON object in model output". Sending `"think": false` on every
/api/chat request fixes it; non-thinking models (qwen2.5) ignore the flag.
Validated end-to-end on deployment hardware (Ollama, qwen3.5:9b): the
grammar checkpoint now caught all five ESL errors in a 3-sentence sample
with correct JSON and string-anchoring, ~8.5s warm.
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