sorokin.oleg.edu/learnharness
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 6 Wiki Activity

fix: verdict death — plan→teach seam inferred from first nodeId-tagged quiz

Browse Source 
The Accept/Decline verdict gate is gone (owner ruling 2026-09-26):
presenting the plan ends the teacher's turn; the learner's next message
is a natural reply. Teaching begins at the first quiz call tagged with a
nodeId (probe quizzes carry none) — both folds (view phaseAt3, native
phaseAt3Lh) advance on the call itself, and the teach divider renders
before that whole assistant block. No backwards compat: verdict markers,
the verdict input event, optimistic appends, composer buttons, and
projection all deleted; prompts and docs rewritten to the natural flow.
This commit is contained in:
amosblomqvist
2026-09-27 00:48:46 +03:00
parent dcb839232c
commit 145b470752
17 changed files with 221 additions and 389 deletions
Download Patch File Download Diff File Expand all files Collapse all files
components/chat.tsx
+13 -54
View File
@@ -128,7 +128,7 @@ const useIsoLayoutEffect = typeof window === "undefined" ? useEffect : useLayout
* as a sub-agent result — rendered as an artifact card, never a bubble. */
/** The quiet system line (custom-message scheme rendering): centered, muted,
* small — the transcript's voice for events that are not speech (wake
* scaffolding, the learner's plan verdict). */
* scaffolding). */
function SystemLine({ children, testId }: { children: React.ReactNode; testId?: string }) {
return (
<div
@@ -402,7 +402,7 @@ export function Chat({
.catch(() => {});
});
const [setupOpen, setSetupOpen] = useState(false);
/** The last explicitly-sent draft (composer send, verdict note): the
/** The last explicitly-sent draft (composer send): the
* refusal flow restores it into the composer and Save replays it.
* Auto-continuations (probe answers) never set it — their refusal
* opens the dialog with no replay. */
@@ -575,35 +575,6 @@ export function Chat({
// derives its own from the same transcript).
const planTree = useMemo(() => latestTree(messages), [messages]);
// ── The plan verdict gate (docs/probe-teach-split.md §4) ──
// The plan has been presented (a tree exists) and the session still sits in
// the planning phase → the composer's send square becomes the stacked
// Accept / Decline buttons. Acceptance is the explicit plan→teach seam
// (a data-plan-accepted part on the message); decline asks for revision.
const planVerdict = useMemo(
() => phaseAt3(messages) === "plan" && planTree !== null,
[messages, planTree],
);
const sendVerdict = useCallback(
(accepted: boolean) => {
const note = text.trim();
if (note) draftRef.current = note; // refusal flow: the note replays as text
setText("");
// TWO REAL MESSAGES, ONE TURN (the custom-message scheme, user-decided
// 2026-09-12): a typed note is the learner's own user message, and the
// verdict is its own custom message — marker part ONLY, no human-
// authored text (the model-facing wording lives in toLlmMessages,
// the native convertToLlm seam; nothing fake is persisted as speech).
// ONE input event, ONE turn: the session hook appends the optimistic
// note + verdict marker (local ids — the settle sync lands the
// server canon) and POSTs {type:"verdict"}; the server persists the
// pair as its own write-ahead — the same single-turn mechanism
// syncFromServer uses, no 409 dance, no double teacher turn.
void send({ type: "verdict", accepted, note: note || undefined });
},
[text, send],
);
// ── Session usage (wave-4 stats bar) ──
// The money channel: the record's turns log, folded client-side by
// core/prices.ts. Seeded once on mount (the page doesn't carry turns),
@@ -1076,21 +1047,17 @@ export function Chat({
// assistant group — is gone). Never a user bubble: the wake is
// scaffolding; the model-facing instruction lives in the adapter.
<PhaseDivider key={m.id || `m-${mi}`} phase="plan" />
) : m.role === "user" && m.parts.some((p) => p.type === "data-verdict") ? (
// The learner's plan verdict (custom-message scheme): the
// accept IS the phase seam — it renders the TEACHING DIVIDER
// (owner ruling 2026-09-26: replaced the "✓ Plan accepted"
// line; same spot, one signal, not two). A DECLINE is not a
// seam — it stays the quiet revision line. Never a user
// bubble: the model-facing wording is the adapter's, and this
// is the learner's ACT, not their words (a typed note, if
// any, renders above as its own real message).
((m.parts.find((p) => p.type === "data-verdict") as
| { data?: { accepted?: boolean } }
| undefined)?.data?.accepted)
? <PhaseDivider key={m.id || `m-${mi}`} phase="teach" />
: <SystemLine key={m.id || `m-${mi}`} testId="verdict-declined">↻ Plan revision requested</SystemLine>
) : (
) : m.role === "assistant" && mi > 0 && phases[mi] === "teach" && phases[mi - 1] !== "teach" ? (
// THE INFERRED TEACH SEAM (owner ruling 2026-09-26, the verdict
// death): teaching begins where the FIRST nodeId-tagged quiz is
// posed — the fold (phaseTimeline, the same derivation the phase
// chip reads) flips to teach exactly AT this assistant group, so
// the divider renders before the whole block (thinking -> prose ->
// the quiz call) — where the explicit accept marker used to sit.
// The boundary check (this index flips, the previous didn't)
// guarantees exactly ONE divider per session.
<PhaseDivider key={`teach-${m.id || `m-${mi}`}`} phase="teach" />
) : (
<Fragment key={m.id || `m-${mi}`}>
<Message from={m.role} className={m.role === "assistant" ? "max-w-full" : undefined}>
<MessageContent className={m.role === "assistant" ? "w-full" : undefined}>
@@ -1234,14 +1201,6 @@ export function Chat({
onStop={abortInFlightTurn}
canSend={canSend}
onEscape={dictation.active ? dictation.cancel : undefined}
verdict={
planVerdict
? {
onAccept: () => sendVerdict(true),
onDecline: () => sendVerdict(false),
}
: undefined
}
/>
}
onSubmitAnswer={(output) => {