Prod autopilot resolved boss/elite fights inline via SimulateCombat, which
swings the enemy once per round (Combatant has no ID to look up the SRD
multiattack profile). Manual !fight uses the turn engine, which loops the
full profile — so autopilot players faced strictly weaker bosses than
manual. D8-e confirmed this is the gap, not a turn-engine artifact.
- Promote the sim's autoResolveCombat/simPickCombatAction to shared plugin
methods autoDriveCombat/pickAutoCombatAction (single source of truth; the
sim now calls the same code prod does).
- Add MessageContext.Silent + a replyDM helper; the turn-engine combat
handlers route their DMs through it so the background autopilot can drive
the real !fight/!attack engine without spamming a DM per round (the EoD
digest summarizes the outcome).
- tryAutoRun now calls runAutopilotWalkDriven (inlineBossCombat flipped
true->false): walk->fight->walk loop so one tick still covers ~autoRunRoomCap
rooms, but boss AND elite now face the player's full kit against the
enemy's full multiattack. Loss surfaces as stopEnded (run already
force-extracted by finishCombatSession).
Trash mobs stay on the fast inline path. GOGOBEE_SIM_INLINE_BOSS=1 A/B
toggle preserved. Build + plugin tests green; sim smoke-run unchanged.
Diagnostic env toggle (off by default) routing the sim's boss/elite
doorways through the inline SimulateCombat path instead of the turn
engine, for A/B-ing the martial T4/T5 'regression'. D8-e confirmed the
gap is honest multiattack math (inline swings the enemy once/round; the
turn engine loops the full SRD profile) and that prod autopilot is
secretly easier than manual !fight. Toggle left in for the D8-f parity
work.
- expedition-sim matrix worker now captures child stderr and dumps
runErr/stderr/stdout-snippet on failure so halted rows have a cause.
- simPickSpiritualWeapon walks slots 2..5 and upcasts when L2 is spent
instead of silently skipping the spell on high-level clerics.
- advanceOnceWithOpts !inlineBossCombat branch now emits the same
"Room X/Y — Boss/Elite. Type !fight to engage." line as foreground.
Adds inlineBossCombat alongside compact in runAutopilotWalk and
advanceOnceWithOpts. Production background autorun keeps both true
(inline auto-resolve), foreground stays both false (manual !fight), the
sim now uses compact=true + inlineBossCombat=false so the boss/elite
doorway returns stopBoss/stopElite after the safety gate — autoResolveCombat
+ simPickCombatAction / simPickSpell drive the fight via the turn-based
engine. The picker (and D8-b upcasting) has been dead since D3's
compact-inline boss rooms; this re-wires it.
n=50/cell L10 smoke vs d7d (zones T1-T3):
bard forest_shadows 61 → 100 (+39)
bard manor_blackspire 10 → 34 (+24)
cleric forest_shadows 15 → 96 (+81)
cleric manor_blackspire 0 → 54 (+54)
fighter T1-T3 100 (unchanged)
Also parallelizes matrix mode via subprocess workers (each child has its
own SQLite — db package globals preclude in-process parallelism). New
-jobs flag, defaults to runtime.NumCPU(). 8 workers gave ~7x speedup on
the smoke matrix.
simPickSpell now enumerates one candidate per available slot >= native
for every prepared damage spell, scored via spellExpectedDamage with the
existing "+1 die per slot" scaling. Cantrips contribute a single slot-0
candidate. Tie-break: highest slot first, then highest expDmg. Picker
returns "<id> --upcast N" when the winning slot exceeds native so
parseCombatCast upcasts in the engine. Build + plugin tests green.
Then: measured zero impact. d8b_corpus.jsonl (same 10x5x100 matrix as
d7d) lands every class within +/-1.5pp of d7d baseline.
Root cause discovered post-corpus: simPickSpell is dead code in the
current sim path. Commit 68ed8e7 ("Long expeditions D3: compact
autopilot auto-resolves boss rooms") added a !compact gate at
dnd_expedition_cmd.go:810 so compact autopilot inline-resolves boss/
elite rooms via resolveCombatRoom -> runZoneCombat -> SimulateCombat
instead of returning stopBoss/stopElite for autoResolveCombat to handle.
The sim uses compact=true (expedition_sim.go:433), so autoResolveCombat
- the only caller of simPickCombatAction/simPickSpell - never fires.
Verified empirically: a stderr-traced simPickSpell produced zero hits
across full bard/cleric L10 runs.
This rewrites the picker-era retrospective. J2's baseline_j2a_v2_all10
(bard 40%, cleric 39%) was measured before D3 with the picker live.
d7d's L10 baseline (bard 34%, cleric 21%) is post-D3 with the picker
disconnected - that 6-18pp drop is the post-D3 caster regression, not
"long-expedition mechanics didn't help casters" as d7d framed it.
D8-a's "+2.2pp cleric" was also noise (1sigma ~ 1.8pp on n=500).
D8-b code kept in tree (correct, currently inert, turns on as soon as
the wire reconnects). Plan rewritten in section 8: D8-b and D8-c are
deferred behind D8-prereq, which splits compact so the sim opts out of
compact-inline boss/elite combat without affecting prod rendering.
Files:
- internal/plugin/expedition_sim.go: simPickSpell upcast enumeration
- gogobee_long_expedition_plan.md: D8-b implemented-but-inert,
D8-prereq added as next step
- sim_results/d8b_corpus.jsonl: 5000-row corpus retained for the
picker-dead baseline
Fix sim regression introduced by D5-b: bare `expedition start <zone>`
returns the loadout prompt DM without outfitting, so SimRunner.RunExpedition
was halting before persisting any expedition. Pass `heavy` from the
harness — tier-max packs, no prod paths touched.
D7-d corpus (sim_results/, gitignored): n=100 × 10 classes × 5 zones ×
L10. Leaderboard mirrors J2b — martials 78–82%, casters 21–42%, cluster
gap unchanged by long-expedition mechanics. Cleric worst at 21% (L10).
Bard/cleric trailers not relieved by autopilot camp pacing; remains
J3-territory. T3/T4 cleared runs hit their §2 target durations.
D5-d retune decision: no change. phase5BDailyBurnRatePct=50 + per-tier
DailyBurn stay as-is — heavy-preset cleared runs end with 78–98% SU
surplus; even TPK runs leave packs mostly full. Supply economy is not
a binding constraint at heavy preset.
Closes the long-expedition track.
cmd/expedition-sim -days N caps runs by synthetic day rollovers
(Outcome="day_capped"). SimResult.DaySnapshots traces HP/SU/threat/rooms
at start, every Night-camp rollover, and end-of-run — unblocks empirical
D5-d retune of phase5BDailyBurnRatePct against per-day SU draws.
maybeAutoCamp / pitchAutopilotCamp / pitchBossSafetyCamp now take a
now time.Time so the sim can inject a synthetic clock; tryAutoRun
still passes time.Now().UTC(). SimRunner.RunExpedition advances simNow
by autoRunCooldown per walk and runs the production camp scheduler
after each soft stop (and pitchBossSafetyCamp on stopBossSafety),
dwelling minAutoCampDwell + breakAutoCampIfDue so the next walk can
proceed. Effect: HP-low mid-day rests, base-camp waypoints, Night-camp
rollovers, and boss-safety holds all fire under the sim; D7-a's
tickEventAnchoredRollover shortcut is retained on TickDay for tests
and the pre-cutoff legacy path.
deliverBriefingEventAnchored reads time.Now().UTC() for its safety-net
check, so synthetic TickDay calls never advanced CurrentDay on D2-b
expeditions — DaysAtEnd / SUEnd stayed at start values. Short-circuit
in TickDay: when isEventAnchored, fire nightRolloverBurn → optional
applyCampRest (Standard, Rough fallback, skipped on low-SU) →
nightRolloverDrift(briefAt). Mirrors pitchAutopilotCamp Night=true.
Production paths untouched.
Unblocks D5-d retune of phase5BDailyBurnRatePct and the class corpus
re-run.
Synthetic sim chars are vanilla base-class (no pet, no subclass), so the
per-round pet attack/deflect/whiff path was never exercised in the sim.
Add -pet-level N (1-10) which stamps a base Massive Dog (no armor) onto the
AdventureCharacter via the normal save path before the run, so combat's
DerivePlayerStats sees HasPet()==true. 0 (default) stays petless.
Measured lift (n=40, 10 classes x L3/7/12 x 3 zones): overall clear-rate
38.3% petless -> 47.8% at pet L10 (+9.4pp); biggest gains go to the caster
trailers (bard +13.3, warlock +11.9, cleric +11.4), narrowing class spread.
Mirror runAutopilotWalk's multi-region auto-advance in the headless sim:
on a mid-zone stopComplete (active multi-region exp with a next region),
advanceToNextRegion and keep simulating instead of scoring a premature
"cleared". A transit error there is now recorded as halted, not cleared.
Also fix misleading run-complete wording: a non-boss region clear of a
multi-region zone now reads "Cleared {region}. The way to {next} opens
ahead." instead of "Cleared {zone}. Run complete." New midZoneRegionClear
helper shares finalizeExpeditionOnZoneClear's gating; the path is shared
with manual !region travel, so both autopilot and manual play get it.
(cherry picked from commit 5d2bba70849a0a3fdeac285cc55ea9b8fadea29c)
autoResolveCombat now calls maybeShortRest after a won fight (HP < 60%
or any slots used, charges > 0), mirroring what a competent prod player
does between rooms. Re-baseline at n=100 across the full 10×3×5 matrix
written to baseline_h5sim_all10.jsonl.
Headline (vs baseline_j2a_v2_all10.jsonl, full delta in h5sim_findings.md):
mage L12 manor +8pp, cleric L12 manor +7pp, but bard L12 manor stays at
1pp and bard/cleric L12 underdark still 0. J2c (defaultKnownSpells
damage option at L3+ for bard, L4+ for cleric) is still the lever for
the trailers. Four borderline -9pp regressions on n=100 cells; flagged
for an n=300 re-run before declaring a sim defect.
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/.
The class-identity audit (98ba416) wired Extra Attack via the new
resolvePlayerSwings helper, but only SimulateCombat (auto-resolve)
called it. The turn-based engine — every !fight/!attack and every
elite/boss gate the sim drives via autoResolveCombat — still called
single-swing resolvePlayerAttack, so Fighter L11+ got 1 swing/turn at
the gates instead of 3. The audit close-out was correct in spirit but
half-applied.
J1 baseline matrix surfaced it: Fighter L12 cleared 100% of T2 forest
but 2% of T3 manor and 7% of T4 underdark, with %boss_reached at 100%
across the board. The wall was the boss-room damage exchange, not
mid-zone attrition. Trace dump on a sample fight: Fighter dealt 79
dmg in 14 rounds (7 hits / 9 swings) — exactly one swing per round —
versus 167 enemy dmg. With multi-swing wired in, the same fight ends
in 7 rounds with the boss dead, Fighter at 87/168 HP, 16 hits in 19
swings.
n=100 matrix after the fix:
Fighter L12 manor: 2% → 100% clr
Fighter L12 underdark: 7% → 98% clr
Fighter L12 forest: 94% → 100% (no leader regression)
Mage cells unchanged (J2 territory). Rogue cells within noise.
Sim infra changes that landed alongside (needed to read the J1
signal):
* expedition_sim auto-arms class-default defensive abilities
(Second Wind / Healing Word) via the new simAutoArmEnabled toggle
+ trySimAutoArm helper, hooked before applyArmedAbility in both
combat builders. Production code paths untouched (toggle stays
off). Without this the sim simulated a player who never types
!arm, which under-counts class survival.
* SimResult.Combats captures per-fight turn-log summaries (rounds,
hits/misses, damage by side, AC values inferred from RollAgainst)
so future J-phase questions can dig into the engine without
re-running the matrix.
* sim_results/run_matrix.sh fans the matrix across (class,level,zone)
cells via xargs -P (one process per cell — each owns its global
sqlite handle). ~6× wall-clock speedup on a 14-core box; n=100
matrix runs in ~3min.
* sim_results/summarize.sh gains p50_yld_clr + %boss_reached columns
so future sweeps don't conflate "reaches boss" with "clears zone".
Baselines:
sim_results/baseline_j0_n100.jsonl — pre-fix (1350 rows)
sim_results/baseline_j1_extra_attack.jsonl — post-fix (4500 rows)
Phase J state: J0 baseline locked, J1 done. T5 dragons_lair still
0% clear universally (J3). Mage T2+ wall still real (J2).
Harden expedition-sim with the three layers a real player carries:
race mods (+1 Human across the board), tier-appropriate equipment
(weapon/armor/helmet/boots/tool promoted from tier 0 to a level-mapped
tier), and a lean consumables bundle (2 heals + 1 buff at T2-T4,
3 heals + 2 buffs at T5). Capture material yields in SimResult so
calibration sweeps actually read out the H4 signal.
Three sweeps in sim_results/ (n=10): bare baseline, +gear/race,
+lean consumables. Headline: yield is gated by combat survival, not
charge counts. Fighter and Mage hit a sharp wall at T3+ even kitted;
no class extracts the T5 dragons_lair boss. Tuning charges before
fixing class survival would over-tune against a strawman, so H4
is now flagged blocked on Phase J.
Phase J appended to gogobee_harvest_charges_plan.md:
J0 baseline at n=30, J1 Fighter T3+ wall, J2 Mage T2+ wall,
J3 T5 boss (sim artifact vs over-tuned vs raid content),
J4 validation matrix.
RunExpedition now drives expeditions end-to-end through the production
plugin paths:
- TickDay calls deliverRecap + deliverBriefing with a synthetic clock
anchored on exp.StartDate, advancing one day per call without
real-time waits.
- Autopilot bursts use compact mode so elite gates auto-resolve inline.
- Boss / surviving elite gates are auto-resolved via handleFightCmd +
looped handleAttackCmd until the session flips Won/Lost/Fled.
- Forks resolve to path 1 deterministically.
- SimResult exposes StopCode, Walks, DayTicks, Threat, plus the full
dnd_expedition_log projected as SimLogEntry rows.
SimRunner re-uses production plugin code paths against a fresh sqlite DB —
SendDM no-ops without a Matrix client, so ground truth comes from the DB.
BuildCharacter persists adv+dnd character rows with class-appropriate
ability scores and spell slots; RunExpedition loops !expedition run with
stall detection until the autopilot halts or the expedition closes.
The cmd binary is a thin orchestrator: one flag-driven run, JSON output.
Per-room structured logging, day-cycle fast-forward, and matrix batching
are next session.
