Solomon Engine — Critical Fixes Plan

Total Fixes

P0 Critical

P1 Serious

P2 Audit

Files Affected

Scoring & Progress Logic

False "All Done!" at Low Scores

P0 Scoring · Backend

▸

Issue

System shows green checkmark and "All Done!" even when score is 2–35%, misleading users into thinking they've completed the interview.

Root Cause

Backend completion threshold was deliberately lowered to ≥ 1% for testing and was never restored. Comment in code even says so.

File

backend/api/routes/interviews.py · line 468

Change

- s.get("segment_completeness_score", 0) >= 1  # Lowered to 1% for testing
+ s.get("segment_completeness_score", 0) >= 70  # Align with frontend threshold

Also aligns with the frontend's existing score >= 70 check in CompletenessResult.tsx:30.

"Continue Anyway" Never Appears on Last Segment

P0 Scoring · Frontend

▸

Issue

Users trapped in "Needs More Detail" retry loop indefinitely on Segment 2 — the escape button is coded to never appear on the last segment.

Root Cause

Condition is retryCount >= 3 && !isLastSegment. The !isLastSegment guard intentionally prevents the button on Segment 2 — the intent was likely to force completion, but instead it creates an inescapable trap.

File

frontend/components/CompletenessResult.tsx · lines 211–221

Change

- {retryCount >= 3 && !isLastSegment && (
+ {retryCount >= 3 && (
    <button onClick={onProceed}>
-     Continue Anyway
+     {isLastSegment ? "Submit Anyway" : "Continue Anyway"}
    </button>
  )}

Skipped Questions Penalize Score

P1 Scoring · Backend

▸

Issue

Skipped answers count as 0 against the full denominator, making it impossible to score 100% if any question was skipped — unfairly punishes valid "skip/pass" responses.

Root Cause

calculate_segment_completeness() evaluates the transcript against all target categories regardless of whether questions were actually asked/answered.

File

backend/services/interview_segments.py · lines 450–530

Change

Track skipped categories in segment data (skip detection already exists in prompts)
Pass skipped_categories as an argument to calculate_segment_completeness()
Filter skipped categories from the denominator before Claude evaluates completeness
Store skipped_categories as a JSON column on the interview_segments table for auditability

Conversation Control

Silence Detection Fires at 11s (Should Be 15s)

P0 Conversation · ElevenLabs Config

▸

Issue

"Are you there?" prompt fires at 11 seconds — too aggressive, interrupts users who are thinking or taking a breath mid-answer.

Root Cause

This is ElevenLabs server-side VAD (Voice Activity Detection), not the frontend 30s/45s constants (those control the auto-end banner only). The ElevenLabs agent has its own inactivity prompt timer set independently on the dashboard.

File

ElevenLabs Agent Dashboard — both Segment 1 & 2 agents

backend/services/interview_segments.py · lines 100, 162

Change

(Dashboard) ElevenLabs agent → Conversation settings → increase inactivity prompt delay from ~11s to 15s for both agents
(System prompt) Add to both segment system prompts: "Do not prompt the user if they pause. Wait at least 15 seconds of silence before checking in."

Zombie Sessions — Auto-End Non-Functional

P0 Conversation · Frontend

▸

Issue

When silence persists for 30–45s the session should auto-end, but it doesn't — creating infinite "zombie" loops the user can't escape.

Root Cause

MIN_CONVERSATION_AGE = 5 * 60 * 1000 (5 minutes) — silence detection is completely disabled for the first 5 minutes of a session. If a session stalls before that mark, auto-end can never trigger.

File

frontend/components/SegmentedVoiceInterview.tsx · lines 38–40, 441–488

Change

- const MIN_CONVERSATION_AGE = 5 * 60 * 1000; // 5 min minimum before auto-end triggers
+ const MIN_CONVERSATION_AGE = 90 * 1000; // 90s minimum — enough to prevent false triggers at start

Also add a WebSocket error handler: if the ElevenLabs connection drops or errors, call endSegment() automatically rather than leaving the user in a blank session.

Premature Termination — Tools & Frameworks Never Reached

P1 Conversation · System Prompts

▸

Issue

Segments end after only 3–4 questions, missing entire categories like "Tools" and "Frameworks" which appear later in the script.

Root Cause

Likely one of two causes: (a) ElevenLabs agent dashboard has a max-turn or max-duration limit set too low, or (b) the SEGMENT_COMPLETE signal in the system prompt is being triggered too aggressively before all categories are covered.

File

backend/services/interview_segments.py · lines 46–173 (both segment prompts)

ElevenLabs Agent Dashboard — check max turn / max duration settings

Change

Add to both system prompts: "You MUST cover ALL of the following topics before signaling completion: [enumerate all categories]. Do not send the completion signal until every topic has been addressed at least once."
Audit ElevenLabs agent dashboard for any max-turn or max-duration settings; increase or remove them
Add a backend guard in the webhook handler: reject SEGMENT_COMPLETE signals if fewer than N categories appear in the transcript

Language Guardrail Broken — Hindi / Tagalog Drift

P1 Conversation · System Prompts

▸

Issue

System ignored Tagalog input, then actively switched to Hindi, and refused to switch back even when the user corrected it in English.

Root Cause

The English-only directive exists in interview_segments.py at lines 100 and 162, but the live ElevenLabs agents on the dashboard may have outdated prompts that predate this guardrail. Code in repo ≠ what's deployed.

File

backend/services/interview_segments.py · lines 100, 162

ElevenLabs Agent Dashboard — verify live system prompts match code

Change

Re-sync both agent system prompts to ElevenLabs dashboard — they may have drifted
Move the English-only rule to the first line of both system prompts (currently buried at the end where the model may deprioritize it)
Strengthen the directive to: "CRITICAL: Conduct this entire interview in English only. If the user responds in any other language, immediately reply in English only. Do not acknowledge, mirror, or switch to their language under any circumstances."

UI / UX Breakdowns

Results Screen Flashes 2s or Fails to Load

P0 UI/UX · Frontend

▸

Issue

After segment completion, the results screen either flashes for ~2 seconds then vanishes, or fails to load entirely, leaving users stranded on a blank screen.

Root Cause

The code has no auto-dismiss timer — but a race condition exists: the component polls for completion, renders the results screen, then a subsequent state update (from a still-running poll interval) transitions the component away from the results view. The poll interval is not stopped when results are shown.

File

frontend/components/SegmentedVoiceInterview.tsx · lines 603+, 792–809

Change

Add a resultsLocked boolean state — set to true the moment completion data is received
When resultsLocked === true, stop all polling intervals immediately (in the state update's effect)
Results view should only unmount when user explicitly clicks "Continue to Next Segment" or "Redo" — never from a background state transition
Add useEffect cleanup on the polling interval that runs when the results view becomes visible

"End Segment" Button Disappears During Transitions

P1 UI/UX · Frontend

▸

Issue

The "End Segment" button appears correctly but vanishes when the AI transitions between topics, trapping users in the session with no exit.

Root Cause

Button requires lastSpokeAtRef.current !== null. During topic transitions the AI resets its state (or reconnects), which may clear this ref — causing the button to disappear even though the AI had already spoken.

File

frontend/components/SegmentedVoiceInterview.tsx · lines 1227–1237

Change

- {elapsedSeconds >= 180 && lastSpokeAtRef.current !== null && (
+ {elapsedSeconds >= 180 && hasAISpoken && (  // hasAISpoken: boolean state, set once, never reset

Replace the ref check with a persistent hasAISpoken boolean state — set to true the first time the AI speaks, and never reset. Unlike a ref, this guarantees the button stays visible through transitions.

#10

No Confirmation Emails on Completion

P1 UI/UX · Backend / Resend Webhook

▸

Issue

Users receive no confirmation email upon interview completion, leaving them uncertain whether their session was recorded.

Root Cause

Two compounding issues: (1) Fix #1 is blocking — the all_complete flag never fires correctly due to the 1% threshold, so the email trigger never runs. (2) Even with Fix #1 applied, the Resend webhook is likely gated on status = 'completed' (100% both segments) rather than any "final" session state.

File

backend/api/routes/interviews.py · webhook handlers

Resend webhook configuration (external)

Change

Dependency: Apply Fix #1 first — this unblocks the completion path
Add an explicit Resend email call directly in the all_complete handler in interviews.py — don't rely solely on the webhook which may have conditions we can't easily debug
Audit Resend webhook config: change trigger from status = 'completed' to status IN ('completed', 'partially_complete') or on any session state that includes all_complete = true

Verification & Audit

#11

Verify Segment 2 Agent Wiring

P2 Audit · Config

▸

Issue

Possible that Segment 2 agent is pulling Segment 1 instructions ("Pain Points") instead of the "Work Style" script.

Root Cause

ELEVENLABS_AGENT_SEGMENT_2 defaults to a hardcoded fallback agent ID (agent_9101kms0edweeyqtgrw2n9n7h2bw). If the environment variable isn't set in production, it silently uses this default — which may not be correctly configured.

File

backend/services/interview_segments.py · line ~120

Production environment variables

Change

Confirm ELEVENLABS_AGENT_SEGMENT_2 is set in production env vars
Open ElevenLabs dashboard and confirm agent agent_9101kms0edweeyqtgrw2n9n7h2bw is configured with the Work Style script
Add a startup log: logger.info(f"Segment 2 agent: {segment_2_agent_id}") for easier future debugging
Raise an error on startup if either agent ID env var is missing (don't silently fall back)

Implementation Summary

#	Fix	Priority	File	Effort
`01`	False "All Done!" at low scores	P0	`routes/interviews.py:468`	S
`02`	"Continue Anyway" missing on last segment	P0	`CompletenessResult.tsx:211`	S
`03`	Skipped questions penalize score	P1	`interview_segments.py:450`	M
`04`	Silence detection fires at 11s	P0	`ElevenLabs Dashboard + prompts`	S
`05`	Zombie sessions — auto-end disabled 5min	P0	`SegmentedVoiceInterview.tsx:41`	S
`06`	Premature termination — misses Tools/Frameworks	P1	`interview_segments.py + Dashboard`	M
`07`	Language guardrail broken (Hindi/Tagalog)	P1	`interview_segments.py:100,162`	S
`08`	Results screen flashes / fails to load	P0	`SegmentedVoiceInterview.tsx:603`	M
`09`	"End Segment" button disappears in transitions	P1	`SegmentedVoiceInterview.tsx:1227`	S
`10`	No confirmation emails on completion	P1	`interviews.py + Resend config`	M
`11`	Verify Segment 2 agent wiring	P2	`interview_segments.py:120 + env vars`	S
11 fixes total		5× P0 · 5× P1 · 1× P2		7× Small · 3× Medium · 0× Large

S < 1 hour M 1–4 hours L 4+ hours