PF2e Combater is a floating combat advisor for Foundry VTT's Pathfinder 2e system. It reads the acting creature's real options — strikes, spells, feats, generic actions — together with the battlefield around it, and helps you plan a whole turn, see it on the canvas, and execute it step by step (with one-click undo). GMs get tactical recommendations for the creatures they run; players plan their own turns.

• Turn planner panel. A floating window for the active combatant showing a draft plan, the remaining action budget (slowed / stunned / quickened aware), and a browsable list of everything the creature can actually do this turn.
• Auto-fill. One click builds a sensible turn from the highest-value actions and fits it to the action economy — moving into reach, attacking, casting, or repositioning as the situation calls for.
• Real action sources. Options are read from the actor itself, not a fixed list:
  • Strikes (melee and ranged), with the multiple attack penalty accounted for — and a per-strike MAP button to pin a level when an ability keeps MAP flat across attacks.
  • Spells — curated catalog plus inferred classification (damage, save, area, control, healing, buff…), respecting slots, focus points, and prepared/spontaneous/innate entries.
  • Generic & skill actions (Stride, Step, Demoralize, Trip, Grapple, Recall Knowledge, …), gated by whether they're usable right now.
  • Class & system feats/actions detected from the actor's items.
• Move-and-strike composites. When a target is a stride or two away, the planner offers combined "Stride → Strike" (and "Stride → Stride → Strike") plays, with the destination computed to land in reach.
• Positional tactics (GM NPCs). Recommends Skirmish/Kite (optional melee → stride out of threat → ranged Strike or offensive spell) and Flank-and-Strike (stride to a flanking square for an off-guard hit) when the situation favors them.
• Situational remedies. A dedicated section surfaces condition responses only when they apply — Stand / Crawl while prone, Retch while sickened, Escape while grabbed/restrained.
• Weapon & position handling. Draw, Sheathe, Release (drop), and Reload for the right weapons; a 1-action Drop Prone when it helps.
• GM aggro targeting. For GM-run NPCs, the planner weighs which player character is the best target (healer, caster, controller, defender, low HP) and pre-picks it.

• Movement preview for strides: the path, reachable squares, the starting square, and the landing square — all clamped to the creature's actual Speed and the scene's diagonal rule.
• Choose the movement type. When a creature has more than one Speed, a Stride can travel on fly, burrow, swim, or climb — the chosen speed sizes the reachable range and is the movement the token uses on execute (land-only creatures see no extra control).
• Vertical movement for fly and burrow: hold Shift and scroll while placing a waypoint to raise or lower its elevation. Each waypoint keeps its own height (so a path can climb, level off, then dive), the elevation is shown on the waypoint, and every leg's vertical distance counts against Speed.
• Teleport destinations. Translocate and other teleports show a range ring and let you click the target space; executing casts the spell and repositions the token instantly, and reverting returns it to where it started.
• Range and area overlays when hovering ranged or template spells, so you can see where a spell reaches before committing.
• Area templates with auto-expiry. Placing a lingering area (cloud, wall, darkness) creates a linked PF2e countdown effect on the caster and removes the template automatically when the duration ends — in encounter or exploration.

• Run the plan step by step, or each step on its own. Strikes read your current target; moves go to the planned destination; spells cast through the system.
• Revert a completed step's real effect — movement returns the token, Stand re-applies prone, area actions delete their template — and Reset undoes the whole turn in reverse. Effects that can't be safely auto-undone (e.g. a condition applied to another token) are flagged for manual cleanup.

• Players plan and execute their own combatant; the GM window follows the selected token.
• Players can share their draft with the GM, who can execute it on their behalf if they go AFK.
• If a player is not currently connected (e.g. their PC is being run by the GM for the session), the GM gets full Auto-fill/edit/execute rights on that character — same as any NPC. The read-only "mirror the player's live plan" mode only applies while the owning player is actually online.

🚧 Work in progress. The Auto-fill planner and the GM aggro/target-picking engine are actively evolving. The exact scores, weights, and heuristics below are current-but-not-final and will change as they're tuned — treat the numbers as illustrative, not a contract. Feedback, disagreements, and "it picked a weird turn here" reports are genuinely wanted: please open an issue with the situation.

No AI, no LLM, no cloud. Auto-fill is 100% deterministic, rules-based math that runs entirely in your browser. Nothing is sent anywhere, it works offline, and the same board always produces the same plan. It reads the PF2e rules and the battlefield — it doesn't "ask a model."

Auto-fill isn't a fixed script — it scores every action the creature can take, then searches for the best combination that fits the turn. Two stages:

Every option starts from a base value by how confidently it's understood (curated spell > strike > generic action), then gets adjusted for the actual situation:

• Disqualified outright (never suggested) when there's no valid target, the target is out of range, the target is immune to the effect, or the action would be redundant (a buff the target already has, an aura that's already up, an untrained skill).
• Rewarded for expected impact: a Strike that's in range and hits hard, and — for the NPCs a GM runs — damage that lands on a weakness, a save spell weighed by the target's actual save odds against the DC, and hitting the most valuable target (low HP, healer, caster, whoever's threatening it). Players never get those hidden-knowledge factors — see No metagaming below.
• Penalized for waste: repositioning to a target already in reach, firing a volley weapon point-blank, or re-applying a buff someone already has.

• Works out your real action budget — 3 actions, minus Slowed / Stunned / already-spent, plus one if you're Quickened.
• Explores legal combinations of the highest-scoring actions and throws out the nonsensical ones: two moves in a row, Drop Prone → Stride, a Step you don't need because the attack already reaches, more than two Strikes, or using a Quickened action on something other than Strike/Stride.
• Accounts for the multiple attack penalty as attacks stack up, and prefers plans that use your whole turn rather than leaving actions on the table.
• Orders the result sensibly — setup before payoff (Demoralize/Feint first, Stand before you move).

Everything that depends on hidden knowledge — a target's weaknesses, resistances, immunities, save DCs, AC, skill DCs, hidden traits, and "which PC is the juiciest target" — is GM-only. When a player auto-fills, the enemy's defensive data is never even loaded, so none of those factors touch the score, and any explanation text that would leak them is stripped out too.

That splits the scoring into two honest modes:

• Player auto-fill ranks actions on what the player legitimately knows: the action's base value, whether the target is in range, the player's own average damage, the multiple attack penalty, movement, action economy, and the player's own conditions.
• GM auto-fill (for the NPCs the GM runs) adds the full defense-aware math below — the GM already knows their monsters' stats, so there's no metagaming.

Every action the creature has lands in exactly one "source" bucket, and the bucket sets the starting score. The bucket reflects how confidently Combater understands the action, not how good it is — it's just a small tie-breaking prior. The situational math below (a Strike in range is already +24, damage adds up to +40, etc.) easily overturns it, and a great-fit action from a low bucket beats a poor-fit one from a high bucket every time.

So there's no spell or action that gets ignored — unrecognized ones simply start at 20 and rise or fall on their situational fit like everything else.

Beyond the base, a few of the actual values the engine uses (🔒 = GM-only, skipped for player auto-fill):

Hard limits: max 2 Strike steps per plan, an action budget of 3 (± Slowed / Stunned / Quickened), and a Quickened action may only be a Strike or Stride.

When the actor is hurt (own HP < 50%), self-preservation factors climb — healing, defensive, and retreat options gain extra weight — so a badly wounded creature is nudged toward staying alive rather than trading blows.

Everything above scores actions. The aggro engine answers the other half — who to point them at — and it's GM-only, for the NPCs a GM runs. When a player auto-fills, aggro contributes nothing (it would require reading party members' kits), so this whole layer is skipped.

For each possible target, the engine builds an aggro profile by reading that creature's kit — the slugs, names, and traits of its spells, feats, and actions (identity fields only, not the rules prose, so a description that merely mentions a condition doesn't trip a false match) — and matching them against role cue-words. Each role a target fills adds to its priority:

Then the target's priority is weighted by the action you'd use on it, so the engine spends the right tool on the right enemy:

• Caster / controller → control, debuff, and grab actions are worth ~1.7–1.8×, plain damage only ~0.8× — shut the dangerous ones down rather than chip their HP.
• Finisher target → control/debuff ×4, everything else ×1.15 — lock in the kill.
• Main defender (the tank) → almost everything is discouraged (×−0.45), only control ×0.35 — don't waste swings on the wall built to eat them.
• Immediate threat → a mild damper (×0.65) so the NPC isn't tunnel-visioned on whoever's adjacent.

The finisher and immediate-threat cues need no hidden data, but because the whole layer only runs for GM-controlled NPCs, players never see or benefit from any of it.

"Swing twice, or hit and run?" — Fighter with a non-agile weapon (avg 11 damage), one goblin in reach, three actions.

• Strike #1 ≈ 46 + 24 + 22 = 92
• Strike #2 = 92 − 15 (MAP) = 77
• A third Strike is blocked (2-Strike cap), so the last action goes to the next-best legal option — Raise a Shield, Demoralize the goblin, or Stride to a second enemy.
• Result: Strike → Strike → Demoralize, not three flailing swings.

"Which damage type?" 🔒 (GM-run NPC) — an enemy caster with both a fire and a cold spell, targeting a PC with fire weakness 5.

• Fire spell: +min(5 × 4, 45) = +20
• Cold spell (no weakness): +0 → the fire spell is picked. If the PC were fire-immune, the fire spell takes −70 and drops out of consideration entirely.
• When that PC plans their own turn, they get no such adjustment against enemies — the planner never reads a foe's weaknesses for a player.

"Is the save spell worth it?" 🔒 (GM-run NPC) — an enemy caster's Fireball (basic Reflex) at a PC. The engine rolls the PC's save against the spell DC across all 20 die faces and weights the outcomes: multiplier = P(crit fail) × 2 + P(fail) × 1 + P(success) × 0.5. A weak-save target pushes the multiplier up (bigger odds bonus and bigger expected-damage bonus); a strong-save target drags it below a plain Strike. For a player, this whole comparison is skipped — the save spell is scored on its base value alone.

"Is Demoralize worth an action?" 🔒 (GM-run NPC) — an NPC deciding between Demoralize (Intimidation) on two PCs. The engine rolls the Intimidation modifier against each PC's Will DC across all 20 die faces and weights the degrees of success — a critical success (frightened 2) counts extra, a critical failure counts against it. Against a low-Will PC that means frequent successes and the occasional crit, so Demoralize scores well; against a high-Will PC it slides toward failure and the NPC leans on the softer target. An untrained NPC also eats −6 (or the action is hidden outright). Proficiency rank and skill modifier both feed this — rank gates/penalizes, the modifier drives the odds. For a player, the target's DC is never read, so a skill action is ranked on its base value and own conditions only.

"Buffs, but only useful ones." — Heroism on a martial ally: 12 (ally) + 24 (attack buff on a martial) = +36. The same ally who already has Heroism: −60 → never suggested. A wounded ally under fire nudges it up a further +14. (This one works for players too — it reads ally class and buff state, not hidden enemy stats.)

"Heal now, or keep swinging?" — a cleric with a Heal spell. If nobody is below half HP, the Heal takes −10 and a Strike wins. The moment the cleric or an ally drops under 50%, Heal gains +34 and jumps ahead — and if a party member is dying, the healing role is prioritized outright. (Reads ally HP, not hidden enemy stats, so it works for players too.)

"Fireball placement." 🔒 (GM-run NPC) — a caster's Fireball. Two PCs clustered with no allies nearby: 34 + 18 × 2 = +70, plus the Reflex-save math. Drag the same blast so it also clips an NPC ally: each ally in the radius is −18, quickly dropping it below a single-target spell — so the engine prefers the clean placement.

"Who do I hit?" 🔒 (GM-run NPC) — a monster faces a fighter (AC 26, sword) and an enemy cleric (spells, a Heal). The cleric profiles as healer + caster (42 + ~20), the fighter as main-defender + main-attacker. A damaging Strike aimed at the cleric keeps most of that priority, but aimed at the tank it's multiplied by −0.45 — so the engine steers the attack at the healer. Had the monster a Slow spell, it'd weight even harder toward the caster (control ×1.8). Whenever a PC drops to dying, that PC lights up as a finisher target and jumps the queue.

"Don't waste the spare action." — Enemy already in reach with one action to spare: a generic Stride toward it is forced to −10, which is worse than the −1 for simply leaving the action unused — so the planner leaves it empty rather than tack on a pointless move. (Still addable by hand in Browse.)

The highest-scoring plan is what lands in the panel. It's always a suggestion — every step is editable, and Browse lets you build the turn by hand instead.