mocks/cloudflare.ts (1)

964-977: hasPrompt has type false | number, not boolean — nitpick.

typeof promptRaw === 'string' && promptRaw.trim().length short-circuits to false or returns the length number. Functionally correct (0 is falsy), but naming implies boolean.

🔧 Proposed fix

-			const hasPrompt = typeof promptRaw === 'string' && promptRaw.trim().length
+			const hasPrompt = typeof promptRaw === 'string' && promptRaw.trim().length > 0

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@mocks/cloudflare.ts` around lines 964 - 977, The variable hasPrompt currently
evaluates to either false or a number because it uses "typeof promptRaw ===
'string' && promptRaw.trim().length"; change it to a boolean expression so the
name matches its type — locate promptRaw/hasPrompt in the mocks/cloudflare.ts
snippet and replace the right-hand side with an explicit boolean check (e.g.,
promptRaw.trim().length > 0 or Boolean(promptRaw.trim().length)) so hasPrompt is
strictly a boolean; keep the surrounding logic that checks hasMessages ||
hasPrompt and return the same jsonOk payload unchanged.

app/utils/call-kent-episode-draft.server.ts (1)

24-33: id and user are selected from call but never used.

draft.call.id and draft.call.user are not referenced anywhere in the function body; only base64, title, and notes are consumed.

🔧 Proposed fix

 			call: {
 				select: {
-					id: true,
 					title: true,
 					notes: true,
 					base64: true,
-					user: { select: { id: true } },
 				},
 			},

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/utils/call-kent-episode-draft.server.ts` around lines 24 - 33, The
include for call is selecting call.id and call.user but those properties
(draft.call.id and draft.call.user) are never used in the function; remove id
and user from the select inside the include (leave base64, title, notes) to
avoid fetching unused data, or if you intended to use them, reference
draft.call.id and draft.call.user where needed—update the include in
app/utils/call-kent-episode-draft.server.ts (the block that constructs the
include for call) accordingly.

app/utils/cloudflare-ai-call-kent-metadata.server.ts (2)

7-26: getCloudflareApiBaseUrl() and getCloudflareWorkersAiAuth() are duplicated from cloudflare-ai-transcription.server.ts.

The transcription utility already defines getCloudflareApiBaseUrl() and uses getRequiredEnv() for auth. Extracting shared helpers to a common module (e.g., cloudflare-ai.server.ts) would eliminate the duplication and the divergence in auth strategies.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/utils/cloudflare-ai-call-kent-metadata.server.ts` around lines 7 - 26,
Extract the duplicated helpers getCloudflareApiBaseUrl and
getCloudflareWorkersAiAuth into a shared module (e.g., cloudflare-ai.server.ts),
export getCloudflareApiBaseUrl and a unified auth helper that uses the existing
getRequiredEnv strategy used by cloudflare-ai-transcription.server.ts, and
update both cloudflare-ai-call-kent-metadata.server.ts and
cloudflare-ai-transcription.server.ts to import these shared functions instead
of defining them locally; ensure the shared auth helper preserves the local
MOCKS=true behavior (return mock values and a MOCK_ token prefix) and otherwise
uses getRequiredEnv to fetch CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_TOKEN.

---

28-36: isCloudflareCallKentMetadataConfigured() returns false when the generator would actually succeed.

generateCallKentEpisodeMetadataWithWorkersAi falls back to a hardcoded '@cf/meta/llama-3.1-8b-instruct' default when neither model env var is set, but isConfigured() requires at least one of them to be truthy. Any route handler that gates "Generate AI metadata" UI on isConfigured() will wrongly hide the feature whenever both vars are absent — even with MOCKS=true.

🔧 Proposed fix

 export function isCloudflareCallKentMetadataConfigured() {
 	const { accountId, apiToken } = getCloudflareWorkersAiAuth()
-	return Boolean(
-		accountId &&
-			apiToken &&
-			(process.env.CLOUDFLARE_AI_CALL_KENT_METADATA_MODEL ||
-				process.env.CLOUDFLARE_AI_TEXT_MODEL),
-	)
+	// A model always resolves (hardcoded default in the generator), so only
+	// account credentials gate availability.
+	return Boolean(accountId && apiToken)
 }

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/utils/cloudflare-ai-call-kent-metadata.server.ts` around lines 28 - 36,
The current isCloudflareCallKentMetadataConfigured() incorrectly requires at
least one of CLOUDFLARE_AI_CALL_KENT_METADATA_MODEL or CLOUDFLARE_AI_TEXT_MODEL
to be set, but generateCallKentEpisodeMetadataWithWorkersAi can succeed using
its hardcoded default or with mocks; update the check in
isCloudflareCallKentMetadataConfigured() (which uses
getCloudflareWorkersAiAuth()) to return true if credentials are present OR if
process.env.MOCKS === 'true' (and don’t require the model env vars), so the UI
gating reflects actual ability to run
generateCallKentEpisodeMetadataWithWorkersAi.

app/routes/me/_layout.tsx (1)

107-116: callTitle and callNotes are selected but never consumed.

Neither entry.callTitle nor entry.callNotes is referenced in the callKentCallerEpisodesDisplay mapping (lines 236–266), so these fields are fetched for nothing.

🔧 Proposed fix

 			select: {
 				id: true,
 				transistorEpisodeId: true,
 				isAnonymous: true,
-				callTitle: true,
-				callNotes: true,
 				createdAt: true,
 			},

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/routes/me/_layout.tsx` around lines 107 - 116, The query is selecting
callTitle and callNotes but those fields are never used in the rendering
function callKentCallerEpisodesDisplay; remove callTitle and callNotes from the
select block (or alternatively use them inside callKentCallerEpisodesDisplay
where entry is mapped) so we don't fetch unused data—update the select object
that contains id, transistorEpisodeId, isAnonymous, callTitle, callNotes,
createdAt to omit callTitle and callNotes (or add references to
entry.callTitle/entry.callNotes inside callKentCallerEpisodesDisplay) to resolve
the unused-field issue.

app/routes/calls_.record/new.tsx (1)

148-161: Prefill correctly uses notes in place of description.

Removal of the AI disclosure prefix from the prefill data is consistent with the broader notes-based workflow.

One small simplification: cleanedQuestion ? cleanedQuestion : '' is always a string (from .trim()), so you can just use cleanedQuestion.

Optional simplification

-									notes: cleanedQuestion ? cleanedQuestion : '',
+									notes: cleanedQuestion,

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/routes/calls_.record/new.tsx` around lines 148 - 161, The onAcceptAudio
handler in CallKentTextToSpeech sets prefill using a redundant conditional for
notes; replace the ternary expression setPrefill({... notes: cleanedQuestion ?
cleanedQuestion : '' ...}) with notes: cleanedQuestion (since cleanedQuestion is
already a string after .trim()), keeping the rest of the handler (setAudio,
setPrefill, scrollToRouteTop) unchanged to preserve the notes-based workflow and
removal of the AI disclosure prefix.

prisma/schema.prisma (2)

123-128: Redundant @@index([callId]) — the @unique constraint already creates an index.

Line 124 declares callId String @unique``, which implicitly creates a unique index on callId. The explicit `@@index([callId])` on line 127 is therefore redundant and can be removed.

♻️ Suggested fix

   @@index([status, updatedAt])
-  @@index([callId])
 }

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@prisma/schema.prisma` around lines 123 - 128, Remove the redundant explicit
index on callId: the field declaration callId String `@unique` already creates a
unique index, so delete the @@index([callId]) line; keep the existing relation
(call Call `@relation`(...)) and other indexes such as @@index([status,
updatedAt]) untouched.

---

113-115: Storing full episode audio as base64 in the database can bloat the DB significantly.

A stitched MP3 episode encoded as base64 can easily be 10–40 MB per draft row. While this works for SQLite and the data is temporary (deleted after publish), it will increase DB size, backup times, and WAL pressure. Consider storing the audio in a file/blob store (e.g., local disk or S3) with a path reference in the DB instead.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@prisma/schema.prisma` around lines 113 - 115, The schema currently stores
large base64 MP3 blobs in episodeBase64 (prisma/schema.prisma), which will bloat
the DB; instead create a small string column (e.g., episodePath or episodeUrl)
to store a filesystem/S3 object key and move actual MP3 data to a blob store
(local disk, S3, or similar) via your upload path logic (store/upload when
generating the stitched mp3 and delete/replace on publish/delete). Update any
code that reads/writes episodeBase64 to use the new storage service and the
episodePath/episodeUrl field (and implement migration to copy existing base64
blobs out to the chosen storage and populate the new path field, then drop
episodeBase64).

app/routes/resources/calls/save.tsx (4)

776-800: Consider adding a catch-all type for all supported intents.

The RecordingIntent type (line 47) only covers 'create-call' | 'delete-call', but the action handler also dispatches 'create-episode-draft', 'undo-episode-draft', 'update-episode-draft', and 'publish-episode-draft'. While RecordingIntent is scoped to the public form, having a union type for all action intents would improve type safety and documentation:

type ActionIntent =
  | 'create-call'
  | 'delete-call'
  | 'create-episode-draft'
  | 'undo-episode-draft'
  | 'update-episode-draft'
  | 'publish-episode-draft'

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/routes/resources/calls/save.tsx` around lines 776 - 800, Update the
intent typing to cover every branch handled by the action handler: extend or add
a union type (e.g., ActionIntent) that includes 'create-call' | 'delete-call' |
'create-episode-draft' | 'undo-episode-draft' | 'update-episode-draft' |
'publish-episode-draft', then use that type for the value returned by
getStringFormValue and/or for the intent variable in export async function
action so TypeScript can validate all handled intents in createCall,
createEpisodeDraft, undoEpisodeDraft, updateEpisodeDraft, and publishCall
branches (replace or augment the existing RecordingIntent accordingly).

---

621-636: Non-atomic publish: episode is live on Transistor before local records are finalized.

After createEpisode succeeds (line 587), if callKentCallerEpisode.create (line 623) or call.delete (line 633) fails, the episode is already published on Transistor but the call record persists. A subsequent retry would attempt to re-publish a duplicate episode. Consider:

1. Creating the CallKentCallerEpisode record before the Transistor publish (with a provisional state), or
2. Wrapping lines 623-635 in a Prisma transaction to ensure atomicity of the local DB operations at least.

This is an admin-only flow so the blast radius is small, but worth hardening.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/routes/resources/calls/save.tsx` around lines 621 - 636, The publish flow
is non-atomic: after createEpisode returns, creating the CallKentCallerEpisode
record (prisma.callKentCallerEpisode.create) or deleting the original call
(prisma.call.delete) can fail leaving a live Transistor episode with no matching
completed local state; fix by making the local DB changes atomic — either create
a provisional callKentCallerEpisode record before calling createEpisode (marking
it as pending and update it after success), or wrap the post-publish DB
operations (prisma.callKentCallerEpisode.create and prisma.call.delete) in a
Prisma transaction (prisma.$transaction) so both succeed or both roll back;
update the code paths that reference createEpisode,
prisma.callKentCallerEpisode.create, and prisma.call.delete accordingly.

---

558-568: Draft update before publish may null out fields unintentionally.

When shouldUpdateFromForm is true, the update on lines 559-567 sets any field to null if the trimmed form value is empty. This is intentional for a "save" operation, but during publish, if a single field happens to be submitted as empty (e.g., due to a browser autofill glitch), the corresponding draft field is permanently nulled — even though line 576 will reject the publish. The admin would then need to re-enter the value.

Consider wrapping the draft update in the same validation that gates the publish, or only updating non-empty fields:

♻️ Only update non-empty fields during publish

  if (shouldUpdateFromForm) {
    await prisma.callKentEpisodeDraft.update({
      where: { callId },
      data: {
-       title: formTitle?.trim() || null,
-       description: formDescription?.trim() || null,
-       keywords: formKeywords?.trim() || null,
-       transcript: formTranscript?.trim() || null,
+       ...(formTitle?.trim() ? { title: formTitle.trim() } : {}),
+       ...(formDescription?.trim() ? { description: formDescription.trim() } : {}),
+       ...(formKeywords?.trim() ? { keywords: formKeywords.trim() } : {}),
+       ...(formTranscript?.trim() ? { transcript: formTranscript.trim() } : {}),
      },
    })
  }

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/routes/resources/calls/save.tsx` around lines 558 - 568, The current
draft update inside the shouldUpdateFromForm branch unconditionally sets title,
description, keywords, and transcript to trimmed values or null, which can
unintentionally wipe existing draft fields during a publish attempt; modify the
update logic in the shouldUpdateFromForm branch (the
prisma.callKentEpisodeDraft.update call) to only include fields that are
non-empty after trimming (e.g., only add title, description, keywords,
transcript keys to the data object when formTitle?.trim() !== "" etc.), or
alternatively run the same validation used to gate publish before performing the
update so empty/invalid form values are not written as null; locate references
to shouldUpdateFromForm, formTitle/formDescription/formKeywords/formTranscript,
and prisma.callKentEpisodeDraft.update to implement the conditional data
population.

---

649-690: Fire-and-forget draft processing could leave drafts stuck in PROCESSING.

startCallKentEpisodeDraftProcessing is called with void (line 685), so if the server restarts or the promise rejects unexpectedly, the draft remains in PROCESSING with no recovery mechanism. Consider adding a stale-draft cleanup job or a TTL check in the admin UI that detects drafts stuck in PROCESSING beyond a reasonable timeout.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/routes/resources/calls/save.tsx` around lines 649 - 690, The current
fire-and-forget call to startCallKentEpisodeDraftProcessing(draft.id, ...) can
leave a draft stuck in PROCESSING if the promise rejects or the server restarts;
change the flow so you await or explicitly handle the promise result (wrap
startCallKentEpisodeDraftProcessing in try/catch) and update the
callKentEpisodeDraft record status on failure, and additionally implement a
stale-draft recovery: add a TTL/timestamp column when creating the draft in
callKentEpisodeDraft.create, mark processing start time in
startCallKentEpisodeDraftProcessing, and add a background cleanup job or admin
UI check that finds drafts still in PROCESSING past a timeout and resets or
marks them failed so admins can retry.

app/routes/calls_.admin/$callId.tsx (3)

191-195: Duplicate helper — extract getNavigationPathFromResponse to a shared module.

This function is duplicated verbatim in app/routes/resources/calls/save.tsx (lines 24-28). Consider extracting it to a shared utility (e.g., app/utils/misc.ts or app/utils/navigation.ts).

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/routes/calls_.admin/`$callId.tsx around lines 191 - 195, Extract the
duplicated helper getNavigationPathFromResponse into a shared utility module
(e.g., create app/utils/navigation.ts or app/utils/misc.ts) that exports the
function, then replace the local implementations in both
getNavigationPathFromResponse occurrences with an import from that shared
module; ensure the exported signature and behavior remain identical (accepting a
Response and returning the constructed pathname+search+hash or null) and update
any imports in callers (such as the handlers in routes/calls_.admin/$callId.tsx
and routes/resources/calls/save.tsx) to use the shared utility.

---

208-218: Side effect inside useMemo.

URL.createObjectURL is a side-effectful call. While the cleanup in the useEffect handles revocation correctly, creating object URLs inside useMemo is discouraged because React may discard and re-run memoized computations without running cleanup. Consider using useState with a lazy initializer or useEffect to create the URL:

♻️ Suggested approach

- const audioURL = React.useMemo(() => URL.createObjectURL(audio), [audio])
- const abortControllerRef = React.useRef<AbortController | null>(null)
- const [isSubmitting, setIsSubmitting] = React.useState(false)
- const [error, setError] = React.useState<string | null>(null)
-
- React.useEffect(() => {
-   return () => {
-     URL.revokeObjectURL(audioURL)
-     abortControllerRef.current?.abort()
-   }
- }, [audioURL])
+ const [audioURL, setAudioURL] = React.useState<string | null>(null)
+ const abortControllerRef = React.useRef<AbortController | null>(null)
+ const [isSubmitting, setIsSubmitting] = React.useState(false)
+ const [error, setError] = React.useState<string | null>(null)
+
+ React.useEffect(() => {
+   const url = URL.createObjectURL(audio)
+   setAudioURL(url)
+   return () => {
+     URL.revokeObjectURL(url)
+     abortControllerRef.current?.abort()
+   }
+ }, [audio])

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/routes/calls_.admin/`$callId.tsx around lines 208 - 218, The audio URL is
being created via URL.createObjectURL inside the React.useMemo for audioURL
which is side-effectful; change this to create the object URL in a
React.useEffect (or useState with a lazy initializer) and store it in audioURL
state, then revoke it in the existing cleanup along with
abortControllerRef.current?.abort(); update references to the audioURL variable
(previously from useMemo) to use the new state value and keep the current
cleanup logic in the React.useEffect that revokes the URL.

---

312-318: Consider a fallback for unknown step values.

If the CallKentEpisodeDraftStep enum is extended later, an unrecognized step value will render undefined inside the <H6>. A simple fallback would make this resilient:

- }[step]
+ }[step] ?? 'Processing…'

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/routes/calls_.admin/`$callId.tsx around lines 312 - 318, The stepLabel
lookup can return undefined for unknown CallKentEpisodeDraftStep values; update
the mapping that builds stepLabel (the const stepLabel = { ... }[step]) to
provide a safe fallback (e.g., 'Unknown step' or '') using a default expression
or nullish coalescing so rendering in the H6 always gets a string; locate the
mapping near the variable step and ensure the code uses something like
(mapping[step] ?? 'Unknown step') so future enum additions or unexpected values
don't render undefined.

prisma/migrations/20260222190000_call-kent-draft-episodes/migration.sql (1)

9-9: Operational concern: storing base64 audio as TEXT inline in SQLite may cause performance issues at scale.

An index is an additional data structure that improves the speed of data retrieval operations on a database table at the cost of additional writes and storage space to maintain the index data structure. Beyond indexing, SQLite stores TEXT/BLOB columns inline in the B-tree page by default, so large episodeBase64 values will bloat page reads for all queries on CallKentEpisodeDraft — even those that don't project the column — once rows spill to overflow pages.

If drafts are truly short-lived (deleted after publish), this is probably tolerable for the current traffic. If rows can linger or the table grows large, consider either:

• Moving the audio to an external store (R2/S3) and keeping only a reference URL here, or
• Using a separate CallKentEpisodeDraftAudio table to isolate the large column.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@prisma/migrations/20260222190000_call-kent-draft-episodes/migration.sql` at
line 9, The CallKentEpisodeDraft table’s large episodeBase64 TEXT column will
bloat SQLite B-tree pages; remove or isolate that large payload by either (A)
replacing episodeBase64 with a small reference_url column and storing the actual
audio in external object storage (R2/S3), or (B) creating a new
CallKentEpisodeDraftAudio table that holds the audio (e.g., id, draft_id FK ->
CallKentEpisodeDraft.id, audio BLOB/TEXT, created_at) and delete episodeBase64
from CallKentEpisodeDraft; update any code paths that read/write episodeBase64
to instead write the external URL or use the new CallKentEpisodeDraftAudio
record when producing/consuming draft audio.

app/routes/resources/calls/save.tsx (1)

552-568: Partial form submission can corrupt draft fields in the DB.

shouldUpdateFromForm uses || — so if any form field is non-null, all four fields are sent to Prisma (including null for those absent from the form). The subsequent publish on lines 570–573 falls back to the in-memory draft snapshot, so the current publish succeeds, but if anything fails after the DB update (e.g., Transistor API error at line 587), the DB draft is left with nulled-out fields. The admin then retries into a corrupted draft.

Fix: only include non-null form fields in the data object.

🛡️ Proposed fix

-  const shouldUpdateFromForm =
-    formTitle !== null ||
-    formDescription !== null ||
-    formKeywords !== null ||
-    formTranscript !== null
-
-  if (shouldUpdateFromForm) {
-    await prisma.callKentEpisodeDraft.update({
-      where: { callId },
-      data: {
-        title: formTitle?.trim() || null,
-        description: formDescription?.trim() || null,
-        keywords: formKeywords?.trim() || null,
-        transcript: formTranscript?.trim() || null,
-      },
-    })
-  }
+  const draftPatch: {
+    title?: string | null
+    description?: string | null
+    keywords?: string | null
+    transcript?: string | null
+  } = {}
+  if (formTitle !== null) draftPatch.title = formTitle.trim() || null
+  if (formDescription !== null) draftPatch.description = formDescription.trim() || null
+  if (formKeywords !== null) draftPatch.keywords = formKeywords.trim() || null
+  if (formTranscript !== null) draftPatch.transcript = formTranscript.trim() || null
+
+  if (Object.keys(draftPatch).length > 0) {
+    await prisma.callKentEpisodeDraft.update({ where: { callId }, data: draftPatch })
+  }

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/routes/resources/calls/save.tsx` around lines 552 - 568, The DB update
currently sends all four fields (title, description, keywords, transcript) with
nulls when any single form field is present, corrupting drafts; change the
prisma.callKentEpisodeDraft.update call so its data object only includes the
keys whose corresponding form values are non-null (e.g., check formTitle !==
null before adding title: formTitle.trim(), same for formDescription,
formKeywords, formTranscript), preserving the existing fallback/trim logic and
leaving untouched fields out of the update so partial submissions don't
overwrite stored draft values; keep the shouldUpdateFromForm check but build a
conditional data map and pass that to prisma.callKentEpisodeDraft.update.

app/utils/cloudflare-ai-call-kent-metadata.server.ts (1)

56-68: Consider using unknown instead of any for unwrapWorkersAiText's parameter.

The result: any parameter silences all type-checking inside the function. Typing it as unknown and narrowing explicitly is equally concise here and prevents future callers from silently passing the wrong shape.

♻️ Proposed refactor

-function unwrapWorkersAiText(result: any): string | null {
+function unwrapWorkersAiText(result: unknown): string | null {
   if (!result) return null
   if (typeof result === 'string') return result
-  if (typeof result.response === 'string') return result.response
-  if (typeof result.output === 'string') return result.output
-  if (typeof result.text === 'string') return result.text
-
-  // OpenAI-ish shape (some models / gateways).
-  const choiceContent = result?.choices?.[0]?.message?.content
+  if (typeof result !== 'object') return null
+  const r = result as Record<string, unknown>
+  if (typeof r['response'] === 'string') return r['response']
+  if (typeof r['output'] === 'string') return r['output']
+  if (typeof r['text'] === 'string') return r['text']
+
+  // OpenAI-ish shape (some models / gateways).
+  const choices = (r['choices'] as Array<unknown> | undefined)?.[0]
+  const choiceContent =
+    choices && typeof choices === 'object'
+      ? ((choices as Record<string, unknown>)['message'] as Record<string, unknown> | undefined)?.['content']
+      : undefined
   if (typeof choiceContent === 'string') return choiceContent
-
   return null
 }

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/utils/cloudflare-ai-call-kent-metadata.server.ts` around lines 56 - 68,
Change the parameter type of unwrapWorkersAiText from any to unknown and update
the function to narrow the unknown before accessing properties: keep the
existing typeof checks for string cases, and guard the OpenAI-ish path by
verifying result is an object (e.g., typeof result === 'object' && result !==
null) before using optional chaining into result.choices?.[0]?.message?.content
so no property access assumes the wrong shape; leave the return behavior
unchanged.

app/routes/resources/calls/save.tsx (1)

683-689: Non-atomic deleteMany + create in createEpisodeDraft — wrap in a Prisma transaction.

A concurrent admin request (however unlikely) between deleteMany and create leaves the call with no draft at all. A $transaction makes this atomic:

♻️ Proposed refactor

-  await prisma.callKentEpisodeDraft.deleteMany({ where: { callId } })
-  const draft = await prisma.callKentEpisodeDraft.create({
-    data: {
-      callId,
-    },
-  })
+  const [, draft] = await prisma.$transaction([
+    prisma.callKentEpisodeDraft.deleteMany({ where: { callId } }),
+    prisma.callKentEpisodeDraft.create({ data: { callId } }),
+  ])

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/routes/resources/calls/save.tsx` around lines 683 - 689, The deleteMany
followed by create is non-atomic and can leave a call without a draft if
concurrent requests interleave; update the createEpisodeDraft logic to run both
operations inside a Prisma transaction (use prisma.$transaction) so
callKentEpisodeDraft.deleteMany({ where: { callId } }) and
prisma.callKentEpisodeDraft.create({ data: { callId } }) execute atomically;
locate the code around the existing deleteMany/create calls in
createEpisodeDraft (or where those symbols appear) and replace with a single
transaction that performs the deletion then the creation.

prisma/schema.prisma (2)

126-128: Index on [callId] is redundant with @unique constraint.

The @unique on callId (line 124) already creates a unique index. The explicit @@index([callId]) at line 127 adds a second, non-unique index on the same column. This wastes storage without providing query benefits.

Suggested diff

   @@index([status, updatedAt])
-  @@index([callId])
 }

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@prisma/schema.prisma` around lines 126 - 128, Remove the redundant non-unique
index on callId: since the field has an `@unique` constraint (callId) that already
creates a unique index, delete the explicit @@index([callId]) declaration to
avoid duplicate indexing and wasted storage; update the model by removing the
@@index([callId]) line and leave other indexes (e.g., @@index([status,
updatedAt])) intact.

---

104-128: Consider the operational impact of storing episodeBase64 in SQLite.

The episodeBase64 field will hold full MP3 audio as base64 data URLs (potentially 20–30MB per draft). In SQLite:

• This bloats the database file and backups
• Queries that SELECT * or include this column (e.g., the polling loader) transfer the full blob
• SQLite's single-writer model means large writes during audio persistence block other writes

For production, consider storing the audio in object storage (R2/S3) and keeping only a URL reference in the DB. For MVP, this works — just be aware of the scaling implications.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@prisma/schema.prisma` around lines 104 - 128, The CallKentEpisodeDraft model
is storing full MP3 audio in the episodeBase64 String field which will bloat
SQLite and block writes; instead change the schema to store a lightweight
reference (e.g., episodeUrl or episodeStorageKey on CallKentEpisodeDraft) and
move actual audio persistence to object storage (R2/S3), then update the logic
that writes/reads episodeBase64 (the code paths that create/update
CallKentEpisodeDraft and the polling/loader that SELECTs the draft) to
upload/download audio from object storage and save only the URL/key in the new
DB field.

app/utils/ffmpeg.server.ts (2)

35-71: Full stitch path omits -map — works but is fragile.

In the full stitch branch (line 55), the last acrossfade filter has no named output label, so FFmpeg auto-selects it. This works but is implicit. The fallback branch correctly uses -map '[out]'. Consider naming the final output and adding -map for consistency and resilience against future filter_complex edits.

Suggested diff

 						[a02][response]acrossfade=d=1:c2=nofade[a03];
-						[a03][4]acrossfade=d=1:c1=nofade
+						[a03][4]acrossfade=d=1:c1=nofade[out]
 					`,
+					'-map', '[out]',
 					outputPath,

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/utils/ffmpeg.server.ts` around lines 35 - 71, The full-stitch branch in
the args ternary (when hasStitchAssets is true) leaves the final acrossfade
unnamed so FFmpeg implicitly picks it; update the filter_complex used in that
branch (the string fed to filter_complex in args) to assign a named output label
(e.g., append [,out] to the last acrossfade) and then add the explicit '-map',
'[out]' entries before outputPath—mirror the fallback branch pattern used for
[out] mapping to make hasStitchAssets branch robust; modify the code building
args (refer to hasStitchAssets, filter_complex, args, and outputPath)
accordingly.

---

79-79: Replace deprecated fs.promises.rmdir with fs.promises.rm.

rmdir({ recursive: true }) is deprecated (DEP0147) in Node 16+. Use fs.promises.rm with both recursive: true and force: true options instead.

Suggested diff

-	await fs.promises.rmdir(cacheDir, { recursive: true })
+	await fs.promises.rm(cacheDir, { recursive: true, force: true })

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/utils/ffmpeg.server.ts` at line 79, The code currently calls await
fs.promises.rmdir(cacheDir, { recursive: true }) which uses the deprecated rmdir
API; replace that call with await fs.promises.rm(cacheDir, { recursive: true,
force: true }) so removal is non-deprecated and resilient; locate the invocation
of fs.promises.rmdir that references cacheDir and update the options to include
both recursive: true and force: true.

app/utils/call-kent-episode-draft.server.ts (2)

47-72: Large base64 audio stored in SQLite — consider size implications.

episodeBase64 stores the full stitched MP3 as a base64 data URL. A typical 20-minute episode at 128kbps ≈ ~19MB raw → ~25MB base64, stored in a single SQLite text column. This works but may cause:

• Slow reads when loading the draft (even for status polling)
• DB bloat over time

Consider excluding episodeBase64 from the polling query in the loader (only fetch it when actually rendering the audio preview), or storing it on disk / object storage instead.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/utils/call-kent-episode-draft.server.ts` around lines 47 - 72, The draft
currently writes the full stitched MP3 as a base64 data URL into the
episodeBase64 column (see createEpisodeAudio, bufferToMp3DataUrl and the
prisma.callKentEpisodeDraft.updateMany calls), which will bloat SQLite and slow
polling; change the flow to avoid persisting large base64 blobs: either (A)
persist the episode MP3 to object storage or disk (implement a helper like
persistEpisodeAudio that returns an episodeUrl/filePath) and store only that
reference in the callKentEpisodeDraft record instead of episodeBase64, or (B)
keep storing episodeBase64 but stop returning it in lightweight polling/loader
queries (exclude episodeBase64 from the select used for status polling and only
fetch it when rendering the audio preview). Ensure updateMany writes the
reference field (episodeUrl/filePath) or clears episodeBase64 accordingly and
update any callers that expect episodeBase64 to use the new reference-based
retrieval.

---

74-95: Transcript step uses stale isCloudflareTranscriptionConfigured() check after async work — correct but subtle.

The configuration check at line 83 happens after potentially waiting for audio generation. If configuration changes between starting processing and reaching this step, it would correctly fail. This is fine — just noting the implicit assumption that configuration is stable during processing.

One minor concern: if transcribeMp3WithWorkersAi returns an empty string, it would be stored as the transcript, and step 3 would proceed with an empty transcript for metadata generation. Consider guarding against empty transcripts.

Suggested diff

 			transcript = await transcribeMp3WithWorkersAi({ mp3: episodeMp3 })
+			if (!transcript?.trim()) {
+				throw new Error('Transcription returned empty result')
+			}

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/utils/call-kent-episode-draft.server.ts` around lines 74 - 95, The
transcript step currently allows an empty string from transcribeMp3WithWorkersAi
to be stored and advance processing; add a guard after calling
transcribeMp3WithWorkersAi to treat empty/whitespace transcripts as errors (or
retry) before updating the draft. Specifically, after
transcribeMp3WithWorkersAi(...) and before the
prisma.callKentEpisodeDraft.updateMany that sets transcript and
step='GENERATING_METADATA', validate that transcript is a non-empty string
(e.g., trim length > 0); if invalid, set an errorMessage and/or set step back to
a safe state and return/throw so you don't persist an empty transcript. Keep the
existing isCloudflareTranscriptionConfigured() check as-is but ensure the new
empty-transcript validation prevents advancing with blank data.

app/routes/calls_.admin/$callId.tsx (1)

335-470: DraftEditor form fields lack aria-label or aria-describedby for accessibility.

The form inputs use <label htmlFor> which is good, but the required fields (title, description, keywords, transcript) have no client-side validation feedback or error states. Consider whether the required HTML attribute alone is sufficient given noValidate is not set on this form (it will rely on native browser validation).

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/routes/calls_.admin/`$callId.tsx around lines 335 - 470, The DraftEditor
component's form fields (ids: draft-title, draft-description, draft-keywords,
draft-transcript) lack ARIA references and visible validation feedback; update
DraftEditor to (1) render per-field error message elements (e.g., <span
id="draft-title-error">) for each required input, (2) add aria-describedby on
each input pointing to its error id so screen readers know about errors, (3)
mark error elements with role="alert" and toggle their content/visibility based
on client-side validation state, and (4) wire simple validation logic in
DraftEditor to set/clear these error messages before submit (or set noValidate
on the <Form> and show custom validation) so native required-only behavior is
not the sole feedback mechanism.

app/routes/resources/calls/save.tsx (1)

808-823: updateEpisodeDraft silently ignores attempts to clear a field to an empty string.

Because if (nextTitle), if (nextDescription), etc., are falsy checks, an admin who intentionally blanks a field (e.g., to delete a bad AI-generated keyword list) will see their change silently discarded and the old value retained. Whether this is intentional or a UX gap is worth confirming; if clearing should be supported, the guard should check !== undefined rather than truthiness.

♻️ Proposed approach for explicit-clear support

-  if (nextTitle) updateData.title = nextTitle
-  if (nextDescription) updateData.description = nextDescription
-  if (nextKeywords) updateData.keywords = nextKeywords
-  if (nextTranscript) updateData.transcript = nextTranscript
+  if (nextTitle !== undefined) updateData.title = nextTitle || null
+  if (nextDescription !== undefined) updateData.description = nextDescription || null
+  if (nextKeywords !== undefined) updateData.keywords = nextKeywords || null
+  if (nextTranscript !== undefined) updateData.transcript = nextTranscript || null

(Requires the Prisma model fields to accept null.)

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@app/routes/resources/calls/save.tsx` around lines 808 - 823, The update
builder currently uses truthy checks (nextTitle, nextDescription, nextKeywords,
nextTranscript) so attempts to clear a field to an empty string get ignored;
change those guards to explicit undefined checks (e.g., if (nextTitle !==
undefined) ...) so empty-string values are included in updateData and then saved
via prisma.callKentEpisodeDraft.update; apply the same change for title,
description, keywords, and transcript so deliberate clears are not silently
discarded.