NPC reactions, crowd barks and what NPCs can notice about V

Maturity: Draft. How Cyberpunk 2077's NPCs react to V (stimuli, the reaction manager, crowd fear, look-at and facial reactions), how an NPC picks and plays a generic voice line, which of those lines could plausibly comment on V, what a script can read about V's appearance, V's car and the district, what existing mods already do here, and how a mod could add reactions and text-only barks. Consolidated on 27 September 2026 from the decompiled 2.31 script bundle (final.redscripts SHA-256 2119046f…ee86, redscript-cli 0.5.31), the game's voiceset scenes and English subtitle files serialised with WolvenKit CLI 9.0.1 in a private scratch folder, WolvenKit's generated type definitions (11720772), the Modding Docs at be2f44ee, and four installed mods read in place. Nothing on this page has runtime evidence yet. Grades follow the knowledge rules: source decompiled game scripts, mod scripts or type definitions; resource game resources we serialised; wiki Modding Docs; hypothesis not yet established.

This page answers: can NPCs notice how V looks and what V drives, and react through the game's own reaction, voice and subtitle systems? The product idea and its feasibility note are in the alive ideas register.

Sources

Game-script citations are paths inside the decompiled bundle. Installed-mod citations are paths inside each mod's folder in the mod manager (mods/<mod name>/). Who made each mod and what it taught us is in the community credits.

1. The short answers

• The game already has one appearance reaction. When a crowd NPC does its proximity look-at and V wears nothing in the legs, inner-chest and outer-chest slots, it pulls a disgusted face and plays its "being followed" line instead of a greeting source core/components/scriptComponents/reactionComponent.script:2853-2858, cyberpunk/player/player.script:5404-5409. Everything else about V's look is ignored.
• Crowd voice lines are chosen by trigger name, not by content. GameObject.PlayVoiceOver(npc, n"greeting", …) queues a SoundPlayVo with that name as voContext source core/entity/gameObject.script:743-770. Each generic voice is a .scene whose entry points carry the same names (greeting, bump, fear_foll, …) and lead through a random node to one of about two lines resource. So an NPC can only "say something about V" if its own voice happens to have such a line under a trigger we can fire.
• A few such lines exist. About 15 of the 538 civilian greeting lines comment on V's looks (compliments on style, face or attractiveness; one "you need a new wardrobe"; one "look what the cat dragged in"), each in one specific voice resource. None comments on a car.
• Text-only barks are easy and look native. A script can build a scnDialogLineData (text, speaker, type OverHead, duration) and post it to the UI blackboard; the chatter controller draws it above the NPC like any vanilla overhead line source cyberpunk/UI/subtitles/baseSubtitlesControllers.script:120-212, Audioware r6/scripts/Audioware/Codeware.reds:47-66.
• Everything the assessment needs is readable from script: worn and displayed items with their tags, hidden slots, V's body gender, the observer's archetype visual tags (anim_Posh, anim_Lowlife, …) and affiliation, V's mounted vehicle record (manufacturer, type, UI data) and its damage level, and the current district record source. There is no wealth or class attribute on districts or vehicles source; that has to come from data we derive.
• Two installed mods already do a rule-based version (Street Sense, Responsive NPCs). Both classify clothing through tags or hand-kept item lists and reuse the vanilla greeting, bump and fear_foll triggers with a matching facial reaction source. Their limits (random line content, curated per-mod item lists) are exactly what XF Studio's resolver can improve.

2. The reaction pipeline

flowchart LR
  A["Stimulus: broadcast or direct stim (65 types)"] --> B["ReactionManagerComponent on the NPC"]
  B --> C["Reaction preset rules: stim type to output (LookAt, Bump, WalkAway, Flee ...)"]
  C --> D["Behaviour reaction or crowd fear stage"]
  E["Player proximity (Crowds profile) and LookedAt events"] --> F["Proximity look-at: head/eyes on V"]
  F --> G["TriggerFacialLookAtReaction: facial idle + voice trigger"]
  G --> H["PlayVoiceOver(npc, trigger name)"]
  H --> I["Voiceset scene entry point to random dialog line"]
  I --> J["Audio + lipsync + overhead subtitle"]

2.1 Stimuli and presets source

• A stimulus is a StimuliEvent: source object and position, radius, a Stim_Record, visual or audio propagation, and one of 65 gamedataStimType values (from AimingAt, Bump and CrimeWitness to VehicleHorn, WeaponDisplayed and Whistle) source core/events/stimuliEvents.script:1-60, orphans.script (enum gamedataStimType). The set is a fixed enum, so a mod cannot add a stim type; it can only send existing ones.
• Stims are sent by the StimBroadcasterComponent: TriggerSingleBroadcast, the static BroadcastStim(obj, type, radius) and SendDrirectStimuliToTarget (sic) source core/components/stimBroadcasterComponent.script, reactionComponent.script:3395-3430.
• Each NPC's ReactionManagerComponent maps a stim to a ReactionOutput through its preset's Rule_Records (GetRules, GetReactionOutput) source reactionComponent.script:1519-1600. Outputs are the 24 gamedataOutput values (LookAt, TurnAt, Bump, WalkAway, BackOff, Reprimand, CallPolice, Flee, Panic, …); priorities come from ReactionOutput.<name> records source reactionComponent.script:3441-3447. Presets are typed (Civilian_Neutral, Civilian_Passive, Ganger_Aggressive, Police_Passive, Child, NoReaction, … 27 in all) source.
• A LookAt output makes the NPC grunt (stlh_curious_grunt) and look at the source if it is in front source reactionComponent.script:1664-1680.
• ScriptedReactionSystem only counts fleeing NPCs and picks one police caller source core/systems/reactionSystem.script:2-85; the reaction logic lives per NPC.
• Crowd NPCs keep a fear stage (Relaxed, Stressed, Alarmed, Panic) and their own crowd handling (HandleCrowdReaction, OnCrowdReaction) that plays fear_beg, fear_run and similar triggers source reactionComponent.script:2323-2760.

2.2 Proximity look-at and the facial reaction source

This is the path that matters for "NPCs notice V":

1. When V enters an NPC's Crowds proximity profile, OnPlayerProximityStartEvent checks CanTriggerExpressionLookAt and, if V is in front within 45°, starts a look-at; two seconds later OnProximityLookatEvent calls TriggerFacialLookAtReaction if V is still in front source reactionComponent.script:5210-5262.
2. CanTriggerExpressionLookAt refuses NPCs in combat or alerted, ragdolling, blinded, with NoReaction or Follower presets, bosses, those in a reaction sequence or alarmed/panicking, in lore animations, busy, asleep or in braindance workspots, in dialogue or scenes, or not on autonomous AI source reactionComponent.script:2919-2960.
3. TriggerFacialLookAtReaction picks a facial reaction and a voice trigger source reactionComponent.script:2835-2874:
   • outside a crowd: facial category 3, idle 1, trigger greeting;
   • V armed (or forced fear): category 3, idle 10 (fear), no line;
   • V naked (IsNaked): category 3, idle 7 (disgust), trigger fear_foll, look-at held;
   • otherwise a random personality from nine (SelectFacialEmotion) sets the facial idle, and the trigger is greeting. Vendors never speak here. A MapLookAtVO table that pairs personalities with triggers exists but is not called source reactionComponent.script:3504-3537.
4. The facial reaction is AnimFeature_FacialReaction { category, idle } applied with AnimationControllerComponent.ApplyFeatureToReplicate(npc, n"FacialReaction", …), and reset after a cooldown source reactionComponent.script:2869-2917. The personality table gives the known idles: category 3 idle 1 aggressive, 5 joy or funny, 7 disgust, 8 shock or surprise, 10 fear, 3 sad; category 1 idle 3 curiosity source reactionComponent.script:3456-3502. The clips are the generic NPC emotion idles (facial expressions).
5. Standing too close in front of an NPC repeatedly ("comfort zone") forces the fear face and sends a CrowdIllegalAction stim back to the NPC source reactionComponent.script:3380-3435.

IsNaked reads the gameplay items (GetActiveItem on Legs, OuterChest, InnerChest), not what is drawn source player.script:5404-5409. A V dressed in gameplay slots but showing nothing through the wardrobe is not "naked" to the game, and a wardrobe outfit over empty slots is. What is actually drawn is readable too (§5).

2.3 Look-at from script source

ActivateReactionLookAt builds a LookAtAddEvent targeting V's pla_default_tgt slot with style and optional repeat, duration and upper-body parts; DeactiveLookAt removes it source reactionComponent.script:3539-3670. It refuses in combat and during lore animations. The same event type is how AMM aims a character's eyes and head (poses).

3. How an NPC picks and plays a voice line

flowchart LR
  A["NPC appearance name"] -->|"audioVoiceTagAppearanceMapping (hypothesis)"| B["Voice tag, e.g. civ_low_f_31_enus_15_sml"]
  B -->|"locVoiceTagListResource: voiceTag to voicesetScenePath"| C["Voiceset .scene"]
  D["PlayVoiceOver(npc, n'greeting')"] --> E["SoundPlayVo voContext = greeting"]
  E --> C
  C --> F["Entry point 'greeting' to random node"]
  F --> G["Dialog line: locstring id, lipsync names, overHead style"]
  G --> H["Audio by locstring id, lipsync animation, chatter bubble"]

3.1 Voicesets are scenes resource

• The game ships 529 generic voiceset scenes (sceneCategoryTag: voiceset) under base\open_world\voicesets\ (civilian low, mid, high, homeless, prostitutes and others; gangs; NCPD; security; mercs; a few named characters) resource. Each has one actor, entry points named after triggers, voInfo pairs (inVoTrigger → outVoTrigger such as greeting → greeting_var_1, with durations), and a screenplay store of dialog lines resource.
• A trigger's entry point leads to a scnRandomizerNode (equal weights) whose outputs are section nodes, each holding one scnDialogLineEvent with visualStyle: overHead, a voContext and a voExpression, pointing to a screenplay line (locstringId, male and female lipsync animation names, gender mask) resource.
• The trigger names match the names scripts pass to PlayVoiceOver exactly (greeting, bump, fear_foll, fear_beg, fear_run, stlh_curious, …), so the voice context selects the entry point [resource + source; the native lookup itself is a hypothesis].
• Civilians share a small vocabulary: greeting (522 scenes), bump (511), phone_start/phone/phone_end (501), fear_run/fear_beg/fear_foll (275), the security reprimands rep_* (≈270), stealth stlh_* and combat barks resource. There is no trigger for compliments, stares, outfits or cars.
• Subtitle entries are keyed by string id and have a female and a male variant chosen by V's gender (for example "look at her" / "look at him") resource. Voiced audio and lipsync come from the same string id and the voice's lipsync set (Modding Docs lipsync guide wiki).

3.2 From NPC to voiceset

• locVoiceTagListResource (extension .voicetags) lists locVoiceTag { voiceTag, voicesetScenePath, id, isApuc } [source: WolvenKit types]. Named characters set voiceTag in TweakDB (jackie, panam, …) wiki; crowd Character records in the REDmod TweakDB sources do not resource.
• audioVoiceTagAppearanceMapping lists groups of appearance names with candidate voice tags [source: WolvenKit types]. That this is how crowd NPCs get a voice from their appearance is the likely reading hypothesis. entInjectVoiceTagEvent { voiceTagName, forceInjection } and communityVoiceTagInitializer can set a voice tag at run time or from a community spawn [source: WolvenKit types].
• Quest graphs can play a voiceset trigger directly (questPlayVoiceset_NodeType: puppet, voicesetName, overriding context, expression and visual style) [source: WolvenKit types].

3.3 Lines that could comment on V resource

Census of the civilian greeting entry points (269 civilian voices, 538 lines, 500 distinct texts):

• Each voice's greeting randomiser picks between its two lines, so even the right voice says the compliment only half the time resource.
• The fear_foll lines that the naked reaction plays are "stop following me" lines (tour guide, stop tailing me, I'm broke) resource: they fit only loosely.
• Car comments would need new lines. The voiced route is a consenting voice actor or text only; cloning the original actors' voices is ruled out (alive ideas).

4. Text-only barks and subtitles source

• Subtitle and chatter controllers listen to the UI blackboard UIGameData.ShowDialogLine, HideDialogLine (by CRUID) and HideDialogLineByData, each a variant holding an array source baseSubtitlesControllers.script:120-252.
• A line is scnDialogLineData { id: CRUID, text, type, speaker, speakerName, isPersistent, duration } (an importonly struct scripts can fill) source orphans.script:44905-44926. Line types include Regular, Holocall, OverHead, OverHeadAlwaysVisible, Radio and Narrator source orphans.script:7396-7410.
• The chatter controller draws OverHead lines above the speaker when the speaker is set and is not V or a vehicle, the scene tier is below 3, no main dialogue blocks broadcasts, and V is within 150 m (20 m in tighter scene tiers); OverHeadAlwaysVisible skips those checks. Players can switch chatter off in the subtitle settings source cyberpunk/UI/subtitles/chattersControllers.script:1-10, 175-206.
• Audioware does exactly this for its sounds: it builds the struct with CreateCRUID, posts [line] to ShowDialogLine and hides it by id after the duration source Audioware r6/scripts/Audioware/Codeware.reds:47-66, Callback.reds:1-12. Localised text can come from ArchiveXL localization: onscreens: files, as Responsive NPCs ships source Responsive NPCs archive/pc/mod/responsive_npcs.archive.xl.
• A text bark therefore costs no audio and no voice rights, respects the player's chatter setting, and looks like the vanilla overhead chatter source. Without audio or lipsync the NPC's mouth does not move hypothesis; a facial reaction plus look-at covers most of that.

5. What a script can read

6. What existing mods do

Both reaction mods reuse only vanilla triggers, so what an NPC actually says is whatever its voice's greeting or bump line happens to be source. Street Sense's hand-kept list of mod item records is the per-mod approach XF Studio's architecture contract rules out.

7. How a mod could add reactions

1. Hook the existing moment. Wrap ReactionManagerComponent.TriggerFacialLookAtReaction (or OnProximityLookatEvent) with redscript @wrapMethod, which works on private methods: Responsive NPCs wraps the private Initialiaze and replaces the private CheckStalk and HandleCrowdReaction of the same class source ReactionComponentOverrides.reds:587-1040. The vanilla gating, cooldowns, look-at and reset already apply; the wrapper decides whether V's appearance overrides the vanilla choice [source: the vanilla methods]. A scriptable system could instead poll nearby NPCs, at higher cost and with none of the gating hypothesis.
2. Choose a face and a look. AnimFeature_FacialReaction with the idles in §2.2 and ActivateReactionLookAt with repeat for a stare source.
3. Say something:
   • Text bark (§4): cheapest, no rights questions, localisable source.
   • A vanilla line in the NPC's own voice: add entry points (for example xfs_compliment) to civilian voiceset scenes at Resource/PostLoad, each leading to lines that voice already has and that fit the context (the census in §3.3, extended by tagging lines offline). PlayVoiceOver(npc, n"xfs_compliment") then plays with the right voice and lipsync, and silently does nothing for voices without one [hypothesis: Responsive V proves scene patching at load; whether the voiceset system accepts new entry-point names needs a game test].
   • New voiced lines: a voiceovermap.json registered through ArchiveXL localization: vomaps: ties a .wem to a string id in a scene wiki modding-guides/quest/adding-a-custom-voiceline-to-a-scene.md; Audioware plays arbitrary audio with subtitles source. Both need consenting voices. Extracting and repacking vanilla VO, as some dialogue mods do, redistributes game audio and is not a route for XF mods.

8. Design implications for XF Studio

• The assessment layer should read what is drawn, not gameplay slots: wardrobe, hidden slots and EquipmentEx outfits. The Studio's clothing resolver already knows the drawn garments, their visual tags and what they hide (worn clothing), which gives coverage ("revealing") from data instead of a curated list.
• "Fancy" or "shabby" can come from data the game already has: item quality and tags, the garment's resolved materials, the vehicle manufacturer and UI data. District wealth can be derived from the crowd itself: the share of anim_Posh/anim_MidCorp/anim_Nightlife against anim_Lowlife/anim_Homeless/anim_Junkie NPCs near V, which follows the game's own population data instead of a hand-written district table [hypothesis: needs a sample in game].
• The observer matters as much as V: a posh NPC and a lowlife should read the same outfit differently; anim_* tags, affiliation and preset type are the inputs source.
• Lines from data too: an offline pass could tag all 14,003 distinct voiceset lines by what they could plausibly be about (looks, style, staring, weather, …) with evidence, so the in-game mod ships only a table of (voice, trigger, string id), never audio.

Open questions

1. Does the voiceset system accept a new entry-point name added at Resource/PostLoad, or does it cache entry names natively?
2. Which resource holds the crowd appearance → voice tag mapping, and can a script read an NPC's current voice tag?
3. Does a script-built OverHead line on an NPC show reliably in open world (distance, scene tier, chatter setting), and does it collide with the NPC's own voiced chatter?
4. Where does a vehicle offer name its vehicle, and do all player vehicles have an offer (gifts and quest cars do not appear in shops)?
5. Do anim_* visual tag shares around V actually separate rich and poor districts?
6. Does IsNaked also trigger other vanilla behaviour (for example in scenes or vendors) that an XF reaction should respect?

In-game test asks

Not scheduled; for a later session once a prototype exists: walk into a Japantown crowd (a) naked, (b) with the prototype's "revealing" outfit, (c) in a luxury car and a wrecked one in Kabuki and Charter Hill; capture the chatter bubbles, facial reactions and log lines the prototype writes.

Related pages

NPC direction · Facial expressions · Worn clothing · Photo-mode poses · Runtime access · World and streaming · Alive ideas