mirror of https://github.com/prosolis/gogobee.git synced 2026-07-15 00:32:40 +00:00

Files

prosolis f2c2d774d4 J2: sim picker casts+consumes; T5 raid-content warning

The post-J1 sweep had the casters clustered at 19–22% L12 clear, vs
martials at 70–80%. A per-round trace across 240 boss-room fights showed
why: autoResolveCombat dispatched !attack only — zero spell_casts, zero
mid-fight consumable uses across every caster class. The entire "caster
cliff" was the sim measuring a strawman where casters couldn't cast.

J2a teaches the sim's autoResolveCombat to mirror a competent prod
player: heal at low HP if an inventory consumable is available, otherwise
cast the highest-EV damage spell (slot or cantrip), otherwise swing.
BuildCharacter now seeds the known-spell list via ensureSpellsForCharacter
so the synthetic spellbook is populated. A -trace flag on the cmd
attaches the raw CombatEvent stream to the last combat of each run for
post-hoc diagnostics.

A first re-baseline (n=100, all 10 classes) showed Ranger regressed
-35.8pp — the picker was burning L3 slots on lightning_arrow when
Ranger's weapon chassis (Hunter's Mark + Extra Attack) was the better
play. Added simMartialFirstClass to gate the picker off for Ranger and
Paladin (whose default kit is also weapon-first / no damage spells).
J2c experimented with widening the picker to control + heal spells;
heal-spell preempt cost druid 10pp (slot heals are 10HP vs 40HP
consumables) and control-spell scoring at 22 cost warlock 6.6pp. Both
reverted. Corpora retained under baseline_j2c*.jsonl for the post-mortem
in sim_results/j2b_findings.md.

Post-J2 L12 leaderboard (baseline_j2a_v2_all10.jsonl, n=100):
  fighter 80.0, ranger 80.0, paladin 78.4, rogue 76.8,
  druid 61.6, mage 53.4, sorcerer 50.6, warlock 48.2,
  bard 40.4, cleric 39.0.

The caster cluster is dissolved; martials are within ±5pp of J1 (sweep
noise). Bard/cleric still trail, but it's no longer a sim artifact —
their defaultKnownSpells damage rosters cap at L2 and the picker can't
pick spells they don't have. That's a prod-level fix, deferred.

J3 trace (sim_results/j3_findings.md): T5 dragons_lair walls every solo
class at 0% (Infernax 546 HP vs solo player HP 110–175; ~25% boss HP
eaten before TPK across all classes). Per the J3 plan menu, this is
party-shaped content the engine doesn't yet have parties for. Surface
a TwinBee-voiced heads-up in handleDnDExpeditionCmd's start path and a
matching tag in !expedition list — players see "raid-shaped — solo
runs not yet survivable" before they spend outfitting coin. No combat
or class balance changes.

Files: cmd/expedition-sim/main.go +trace flag; expedition_sim.go picker
+ SimCombatSummary.Events + spellbook seed; dnd_expedition_cmd.go
raidContentWarning + list tag. All baselines + traces + findings
checked in under sim_results/.

2026-05-17 15:43:41 -07:00

6.3 KiB

Raw Blame History

J2 — Caster boss-survival traces (per-round event sweep)

Date: 2026-05-17 Corpus: sim_results/j2_traces.jsonl — 240 expeditions Cells: {mage, cleric, sorcerer, warlock, bard, druid} × {manor_blackspire, underdark} × L12 × n=20 Method: New -trace flag on cmd/expedition-sim attaches the raw CombatEvent stream to the last combat (the boss/elite room) of every expedition. Aggregated in /tmp/analyze_j2.py.

Headline

The caster cliff in baseline_j1_all10.jsonl is not a class-balance problem. It is a sim-artifact: autoResolveCombat dispatches !attack only — never !cast, never !consume. Across 240 boss-room fights:

Event type	Count
`player` `spell_cast`	0
`player` `use_consumable`	0
`consumable` events	0
`player` `hit` (weapon swing)	381
`player` `thorn_lash` (druid passive)	261

Every "caster boss-fight" the matrix has scored to date is a melee swing-fest with the caster's spell slots untouched and their consumables sitting in inventory.

Class-level per-cell boss-fight rollup (L12)

class	zone	n	reachBoss	lost	rnd	pDmg	eDmg	noSlot	eHPleft
mage	manor_blackspire	20	10	10	7.4	27.4	89	10	116.6
mage	underdark	20	13	13	5.4	28.7	68.1	13	133.3
cleric	manor_blackspire	20	7	7	7.6	28.1	101	7	115.9
cleric	underdark	20	1	1	2	12	26	1	150
sorcerer	manor_blackspire	20	11	11	6.8	16.9	86.2	11	127.1
sorcerer	underdark	20	6	6	6.5	26.8	79	6	135.2
warlock	manor_blackspire	20	10	10	8.8	29.9	108.4	10	114.1
warlock	underdark	20	15	15	8.6	53.4	96.1	15	108.6
bard	manor_blackspire	20	14	14	9.1	29.9	103.8	14	114.1
bard	underdark	20	10	10	7.8	46.9	76.5	10	115.1
druid	manor_blackspire	20	15	15	8.1	77.8	107.5	15	66.2
druid	underdark	20	14	14	6.5	72.3	81.1	14	89.7

Columns: slotRuns = boss fights where ≥1 slot spell fired; casts = total slot casts; noSlot = boss fights with zero casts; eHPleft = mean enemy HP remaining at player-loss (the "how close did we come" number).

Why druid is the lone outlier at 39% mean clear

Druid's ThornLashDmg = 2 + level/4 passive (dnd_passives.go:210, Class-identity audit 2026-05-16) fires every time the enemy lands a hit on the player — 5 reactive damage at L12, ×~10–15 enemy hits per fight = 50–75 incidental player damage. That alone accounts for the ~70+ mean pDmg druid posts vs 17–53 for the trailers. Druid is not "tankier than the trailers" — druid is the only class whose damage profile doesn't depend on the spell action. Wild Shape, the actual durability beat the plan called out, is not wired through autoResolveCombat at all; thorn lash is.

Reaction-spell hypothesis (plan §6.J2 H2) is moot for now

dnd_spells.go already documents that reaction-cast spells (shield, counterspell, hellish_rebuke) are explicitly skipped today — no reaction window in combat. So even if autoResolveCombat were upgraded, Shield-style auto-fires aren't a J2 lever until a reaction phase exists.

Conclusion — J2 is hypothesis-3-shaped but the cause is upstream

The plan listed four hypotheses; the trace evidence speaks to them as follows:

~~Damage falls off vs boss HP~~ — confounded: cantrip damage was never tested; player only swung weapons.
~~No durability backstop~~ — Druid's outperformance is offense (thorn lash), not durability. So the durability gap, if it exists, can't be read from this corpus.
Slot/resource economy — slots aren't being spent in boss fights at all. Confirms casters reach boss "with slots intact" (the plan's optimistic read), but in the worst possible way: the sim refuses to spend them.
Cleric identity — simAutoArmDefaultFor does pre-arm Healing Word for Cleric (dnd_abilities.go:375), and that fires on the armed-ability path before combat. But mid-fight casts of Spiritual Weapon / Bless never fire.

J2a — Teach autoResolveCombat to cast and consume. Before each !attack, the sim should mirror "what a competent prod player would do":

If HP < ~40% MaxHP and a heal-tagged consumable is in inventory → dispatch handleConsumeCmd with that item.
Else if the character is a spellcaster with available slots and a known damage/control spell that lands → dispatch handleCastCmd (priority: damage spell that finishes the enemy → control/stun → buff → cantrip).
Else swing.

The picker can be small and class-blind for the first cut — pick the highest-level slot whose damage roll most efficiently reduces remaining enemy HP, fall back to a cantrip when slots are dry. Mirror combat_bridge.go's SelectConsumables for the heal side.

J2b — Re-run baseline matrix with the fixed autoResolveCombat. The current n=100 baseline measures a strawman. Once J2a lands we get a real read on caster boss-survival. Only then is it worth picking class-level levers (or concluding "casters are fine, the sim was lying").

Out of J2 scope for now: Reaction wiring (Shield, Counterspell). Per dnd_spells.go's own comments, the reaction phase doesn't exist yet; lifting that ban is its own surface.

Side note — `autoCombatRoundCap` vs the actual constant

Plan mentions autoCombatRoundCap = 200; verified in expedition_sim.go:528. No issue, just confirming the harness isn't truncating fights early.

6.3 KiB Raw Blame History Unescape Escape