fix: preserve failed responses when retrying model calls
Keep failed attempts in the saved transcript and omit only trailing failures from Retry's temporary agent history. Use the matching temporary prefix length when persisting genuinely new output. Preserve the error notice across plan/compaction markers, refuse Retry after later compaction, and avoid synthesizing results for tool calls inside failed generations while retaining Pi's native filtering. Verified: 173 focused Retry/session/compaction-projection/practice-record tests, typecheck, changed-file lint, and diff --check. No app launch or real provider calls.
This commit is contained in:
amosblomqvist
+28
-25
@@ -187,6 +187,17 @@ export async function handleChatInput(
|
||||
return Response.json({ error: "Unknown chat" }, { status: 404 });
|
||||
}
|
||||
|
||||
// Refuse an ineligible Retry before catalog refresh, compaction, or any
|
||||
// turn/write effects. Revalidate after the awaited gates below too.
|
||||
if (input.type === "retry") {
|
||||
try {
|
||||
retryDrops(record.messages);
|
||||
} catch (err) {
|
||||
if (err instanceof ChatInputError) return Response.json({ error: err.message }, { status: 400 });
|
||||
throw err;
|
||||
}
|
||||
}
|
||||
|
||||
// THE UNIVERSAL TURN GATE (BYOK, owner, 2026-09-22): a turn needs the
|
||||
// CONVERSATION chain — the chat's own model override over the strong
|
||||
// slot — and there is no factory floor; when both are empty the turn
|
||||
@@ -224,18 +235,6 @@ export async function handleChatInput(
|
||||
return Response.json(envelope, { status: 503 });
|
||||
}
|
||||
|
||||
// THE RETRY GATE (issue #110): a retry with no failure to retry
|
||||
// refuses here, BEFORE compaction and beginTurn — no effects. The
|
||||
// trim itself re-validates on the record compaction re-reads below.
|
||||
if (input.type === "retry") {
|
||||
try {
|
||||
retryDrops(record.messages);
|
||||
} catch (err) {
|
||||
if (err instanceof ChatInputError) return Response.json({ error: err.message }, { status: 400 });
|
||||
throw err;
|
||||
}
|
||||
}
|
||||
|
||||
// THE COMPACTION GATE (owner round 14, 2026-10-05 — the structural
|
||||
// ruling: NOTHING model-facing for this chat interleaves with a
|
||||
// running compaction): await any in-flight compaction (the settle
|
||||
@@ -274,18 +273,17 @@ export async function handleChatInput(
|
||||
const compactedRecord = stores.chatStore.getChat(chatId);
|
||||
if (!compactedRecord) return Response.json({ error: "Unknown chat" }, { status: 404 });
|
||||
record = compactedRecord;
|
||||
// THE RETRY TRIM: recomputed on the current record (the awaits above
|
||||
// let other writers land), still before beginTurn. The trimmed record
|
||||
// reaches disk through the write-ahead.
|
||||
// Recompute Retry eligibility on current truth, still before beginTurn.
|
||||
// A compaction marker or another writer may have landed during awaits.
|
||||
// These omissions affect ONLY the agent context, never the saved record.
|
||||
let retryOmissions = new Set<string>();
|
||||
if (input.type === "retry") {
|
||||
let drop: Set<string>;
|
||||
try {
|
||||
drop = retryDrops(record.messages);
|
||||
retryOmissions = retryDrops(record.messages);
|
||||
} catch (err) {
|
||||
if (err instanceof ChatInputError) return Response.json({ error: err.message }, { status: 400 });
|
||||
throw err;
|
||||
}
|
||||
record = { ...record, messages: record.messages.filter((m) => !drop.has(m.id)) };
|
||||
}
|
||||
|
||||
// The input's native appends (pure; the honest 400 for an answer with no
|
||||
@@ -344,8 +342,13 @@ export async function handleChatInput(
|
||||
// ride this same write-ahead - the pump de-queued them at fire, so this
|
||||
// is the ONLY copy ever appended.)
|
||||
const merged = [...record.messages, ...appends];
|
||||
// Retry preserves every persisted attempt. Pi needs a non-assistant
|
||||
// tail to start continue(), so its temporary history omits the trailing
|
||||
// failures only. Count that SAME temporary prefix at settlement, or the
|
||||
// retained failures would make persistTurn skip genuinely new output.
|
||||
const agentMessages = input.type === "retry" ? merged.filter((m) => !retryOmissions.has(m.id)) : merged;
|
||||
const stagedAutomatic = isSubscriptionProvider(conversation.provider) && input.type === "steers";
|
||||
let persistBaseCount = stagedAutomatic ? record.messages.length : merged.length;
|
||||
let persistBaseCount = stagedAutomatic ? record.messages.length : agentMessages.length;
|
||||
try {
|
||||
if (!stagedAutomatic) stores.chatStore.saveMessages(chatId, merged);
|
||||
} catch (err) {
|
||||
@@ -384,7 +387,7 @@ export async function handleChatInput(
|
||||
// the session is still probing: build the PROBER (quiz/ask + probe_done).
|
||||
// Both share this whole composition (write-ahead persist, liveness,
|
||||
// steers, usage, extraction) — only the agent differs.
|
||||
const teaching = hasProbeDone(merged);
|
||||
const teaching = hasProbeDone(agentMessages);
|
||||
// BYOK (owner, 2026-09-22): the conversation chain was resolved by the
|
||||
// turn gate above (chat.model override → strong slot → refused) — the
|
||||
// engine build is pure from here; thinking levels clamp natively.
|
||||
@@ -403,7 +406,7 @@ export async function handleChatInput(
|
||||
// sub-agent JOB tools (research/demo/visual/pdf; adjudicated with the
|
||||
// user: quiz/ask/fs/plan/probe_done are always-on).
|
||||
if (isSubscriptionProvider(conversation.provider)) requestCycle = new SubscriptionRequestCycle({
|
||||
initialMessages: asAgentMessages(merged),
|
||||
initialMessages: asAgentMessages(agentMessages),
|
||||
staged: stagedAutomatic ? asAgentMessages(appends) : undefined,
|
||||
foregroundOwed: !callPurpose.background,
|
||||
purpose: callPurpose,
|
||||
@@ -419,7 +422,7 @@ export async function handleChatInput(
|
||||
engine,
|
||||
callPurpose,
|
||||
requestCycle,
|
||||
messages: asAgentMessages(merged),
|
||||
messages: asAgentMessages(agentMessages),
|
||||
onUsage: collectUsage,
|
||||
toolToggles: effective.agent.tools,
|
||||
// Debug (owner, 2026-10-04): teacher/prober raw provider events under
|
||||
@@ -515,7 +518,7 @@ export async function handleChatInput(
|
||||
// Debug recorder run marker (owner, 2026-10-04): correlates the
|
||||
// JSONL event stream with the record's turns; no-op unless
|
||||
// LH_PROVIDER_EVENTS=1.
|
||||
providerEventNote(chatId, "turn", `run start (base: ${merged.length} messages)`);
|
||||
providerEventNote(chatId, "turn", `run start (base: ${agentMessages.length} messages)`);
|
||||
const handle = runPiTurn({
|
||||
agent: constructed,
|
||||
requestCycle,
|
||||
@@ -555,8 +558,8 @@ export async function handleChatInput(
|
||||
settleSteers(chatId, persisted, true, constructed);
|
||||
},
|
||||
onFinish: (turnMessages, isAborted) => {
|
||||
// Wave B: the record write is the agent's own final state —
|
||||
// the new messages are the slice past the merged base.
|
||||
// Wave B: append only messages past the temporary agent prefix.
|
||||
// Retry's saved record still includes its earlier failed attempts.
|
||||
const stamped = persistTurn(stores.chatStore, chatId, persistBaseCount, turnMessages, isAborted);
|
||||
noteSteersDurable(chatId, stores.chatStore.getChat(chatId)?.messages ?? stamped);
|
||||
// usage log append (idempotent per turn id) — the stats bar
|
||||
|
||||
Reference in New Issue
Block a user