Bug class — spell-repair LLM function defined but never called, leaving caster characters with "No spell list recorded" until user manually asks "fix my spells". Trigger on /repro when the Spells panel shows the fallback warning OR the LLM-generated spell list is in prose but not in player_character_data.
Bug class — spell-repair LLM function defined but never called, leaving caster characters with "No spell list recorded" until user manually asks "fix my spells". Trigger on /repro when the Spells panel shows the fallback warning OR the LLM-generated spell list is in prose but not in player_character_data.
Spells panel "No spell list recorded" — dead-code autoseed bug class
Symptom
Spells panel shows the fallback warning:
▸ Spell Slots:
• Level 1: 4/4
• Level 2: 3/3
▸ Spells:
No spell list recorded. Type: "What spells do I know?" to set them up.
— for a character with spell_slots populated AND cantrips/spells_known/spells_prepared missing in player_character_data. The panel rendering gate is at $PROJECT_ROOT/main.py:3631-3641:
if (
spell_slots
andnot spells_known
andnot cantrips
andnot spells_prepared
):
lines.append(" No spell list recorded. Type: ...")
Root cause (canonical)
The codebase has three relevant functions in $PROJECT_ROOT/world_logic.py:
Function
Lines
Behavior
_spells_missing_for_class(pc)
6852
Detector — returns matched caster class name or None
_seed_starting_spells_if_missing(pc)
6873
Warn-only — logs ⚠️ Spellcasting class 'X' has missing spells after creation. and returns pc unchanged
_generate_spells_via_llm(pc, user_id)
6978
LLM-backed repair — calls Gemini with the SpellRepairAgent prompt at line 6957, parses JSON, writes cantrips/spells_known/spells_prepared to pc. Wraps the call in BQ logging via _bq_log_spell_repair_interaction. Fail-soft: returns pc unchanged on any error.
_generate_spells_via_llm is defined but never called — grep -rn "_generate_spells_via_llm" $PROJECT_ROOT/ finds only the definition + its BQ log helper, no callers.
The 4 sites that could call it (world_logic.py:4881, :7384, :7596, :7693) all invoke the warn-only _seed_starting_spells_if_missing instead. So the detector fires, the warning is logged, but no repair ever runs.
End state: any caster whose first creation turn doesn't emit spell fields will show the "No spell list recorded" warning forever, unless the user manually types "fix my spells" — which still doesn't invoke the repair function. The LLM just hallucinates the spells because the prompt context now implies it.
Diagnostic recipe
1. Static-evidence greps (run BEFORE asking user for phenotype anchors)
# Confirm dead-code condition
grep -rn "_generate_spells_via_llm"$PROJECT_ROOT/ | wc -l # expect 1 (defn only)
grep -rn "_seed_starting_spells_if_missing"$PROJECT_ROOT/world_logic.py | wc -l # expect 5 (defn + 4 calls)# Confirm the symptom rendering gate
grep -n "No spell list recorded"$PROJECT_ROOT/main.py # expect 1 match near line 3640# Find sibling repros (campaigns with caster + missing spells)
gh issue list --repo $GITHUB_REPOSITORY --state all \
--search "spells panel OR spell list recorded OR fix my spells" \
--json number,title,state
player_character_data.spells_prepared missing or []
3. Verify dead code (the smoking gun)
# In a Python REPL with mvp_site on path:from mvp_site import world_logic
asserthasattr(world_logic, "_generate_spells_via_llm") # existsimport inspect
src = inspect.getsource(world_logic._generate_spells_via_llm)
# Note: no callers anywhere in $PROJECT_ROOT/. The function has been in the codebase# since the spell-repair feature was added but had no caller wired in.
Fix shape (Track A — wiring)
Add a wrapper that calls the LLM-backed repair when a caster has missing spells, falling back to the warn-only path on LLM error:
Replace all 4 call sites of _seed_starting_spells_if_missing in world_logic.py:
:4881 (unified action path — primary entry for character creation completion)
:7384 (God-mode character creation — also write back to player_character_data since the return value is otherwise discarded)
:7596 (Dragon Knight template merge — write back via initial_game_state["player_character_data"] = ...)
:7693 (merged-pc path — write back via initial_game_state["player_character_data"] = ...)
Two of the call sites discard the return value (_seed_starting_spells_if_missing returns pc but the call sites pass pc as the only argument and don't reassign). When switching to _repair_or_warn_missing_spells, you MUST either:
Reassign the local variable: player_character_data = _repair_or_warn_missing_spells(player_character_data, user_id), OR
Write back to the parent dict: initial_game_state["player_character_data"] = _repair_or_warn_missing_spells(_dk_pc, user_id)
Otherwise the LLM-generated spells land in a transient dict that never persists.
Track B (out of scope — file as follow-up)
Prompt-level enforcement: add a hard rule in the character-creation prompts at $PROJECT_ROOT/prompts/ that state_updates.player_character_data MUST include cantrips[]/spells_known[]/spells_prepared[] for any class flagged as a caster. Empty arrays or missing fields will trigger an automatic SpellRepairAgent call.
Track C (out of scope — file as follow-up)
Frontend fallback at $PROJECT_ROOT/frontend_v1/app.js:789 (buildSpellsHTML): when data.spells_summary is non-empty AND data.spells_known/cantrips/spells_prepared are empty, render the prose spells_summary as a fallback list instead of the "No spells available" default. Degradation pattern, not a fix.
Evidence requirements (per AGENTS.md)
Track A modifies $PROJECT_ROOT/world_logic.py (production backend). Per AGENTS.md:
Real-server proof: local Flask server with WORLDAI_DEV_MODE=true + real Firestore creds.
Real LLM capture: the SpellRepairAgent Gemini response that populates spell fields.
Browser/video evidence: headless Chrome recording of the Spells panel transition from "No spell list recorded" → populated list.
BQ event_type=spell_repair row confirms the auto-repair ran.
Same-symptom coverage on a second campaign (twin-copy via scripts/copy_campaign.py).
Deployed dev SHA at repro: aaa03c4d566a160715b170c209aad8c6447b84bb (branch chore/skip-draft-pr-workflow-runs).
Sibling bug classes (NOT this one)
Level-up exit classifier (#7554, #7609) — LEVEL_UP_EXIT_ANCHOR_PHRASES too narrow, freeform "apply recommended ..." not matched. Different file ($PROJECT_ROOT/intent_classifier.py), different mechanism (semantic routing), different fix (anchor coverage).
God-mode directive missing — custom features granted in narrative but never persisted. See references/god-mode-directive-missing-subclasses.md.
NPC status persistence — narrative-only state changes that never reach state_updates.npc_data. See references/npc-status-persistence-bug.md.
The dead-code-autoseed bug is distinct because the fix surface is server-side wiring (existing LLM call), not prompt-only.
Pitfalls
Don't assume the LLM is the bug. The LLM DOES generate correct spells when the user nudges it (Scene 3 of H9rwoizUNH01vpJhVhF4 shows a verbatim SpellRepairAgent-quality spell list in prose). The bug is purely the wiring.
Don't add a regex/keyword guard for "fix my spells". That's a fix to a different (perceptual) symptom; the user should NOT have to know the magic phrase.
Don't change the prompt to hard-require spells for casters only as the primary fix. That works as Track B but doesn't address the existing dead code. Both tracks ship.
Be careful with the 4 call sites: two of them discard the return value. Wiring without reassigning leaves the repair in a transient dict. Use the write-back pattern shown above.
The _generate_spells_via_llm function is fail-soft: it returns pc unchanged on any error. The _repair_or_warn_missing_spells wrapper around it should ALSO be try/except (LLM errors shouldn't crash the turn). On failure, fall back to the warn-only path so we still log the gap.
