AI Roguelite News

[h2]Open-Ended Overhaul, Plausibility Mechanic[/h2]

All open-ended instructions are now categorized into Player Action or Story Manipulation. To circumvent this entire system and use the old open-ended system, use the cheat code opopen 1

[h3]Player Action[/h3]
Encompasses situations where you describe taking any sort of action, whether it be walking somewhere, stabbing a zombie, casting a spell, or anything in between. These will be classified according to one of 3 Plausibility tiers: Trivial, Medium/Hard, and Implausible.

• Trivial: The minimum roll outcome is the same as what you specified under “minimum roll check for basic actions” in Options (default minor failure).
  
• Medium/Hard: Regular roll check with no minimum outcome. An ability may be invoked if it matches the recommended ability, which usually has a higher multiplier than default for things like damage, healing etc.
  
• Implausible: If the action did not require an item from your inventory, an ability will be required. The action will be blocked if the requirement isn’t met. This requirement is bypassed in creative mode and replaced by a roll check penalty.
  

For all Plausibility tiers, an item may be required, which blocks the action if you don’t have it, unless using creative mode.

[h3]Story Manipulation[/h3]

Encompasses situations where you describe events that shouldn’t be within your control, such as specifying what an NPC is doing or describing a change in the environment. Always allowed in creative mode, and costs 1 Essence of Opportunity in regular mode.

[h2]Equipment Overhaul[/h2]

[h3]More slots[/h3]
Added equipment slots for face, neck, inner torso, finger, and pants.

[h3]Better classification[/h3]
Revised the prompt to reduce false negatives for wearable equipment

[h3]Item breakage/depletion mechanic[/h3]
In the old system, an item would deplete only if you used it while it wasn't equipped. Now, it's using the story from the AI to determine whether it should be depleted (depleted in this case meaning broken, expended, given away, stolen, or left your possession in any way). Many items will have a % chance to resist the story-based depletion, which increases by a flat 20% when equipped (therefore most epic and legendary items are indestructible when equipped).

[h3]Use your equipped items[/h3]
In the past, you had to unequip an item to "use" it. Now, you can also use equipped items.

[h3]Revised concept of a "wieldable" item[/h3]
In the past, anything that wasn't a wearable or consumable was deemed "wieldable", reflecting the fact that you can hold almost anything in your hands. Now that using items from the inventory doesn't incur such a large penalty, there's a new category for miscellaneous items, which makes wieldable items rarer and more special.

[h3]Slightly revised armor logic[/h3]
Armor will only be used to "block" an attack in the story prompt if the favorable text triggered. This prevents armor blockage from becoming a bad thing due to the revised story-based breakage mechanic. The formula is largely the same, with each equipped item having an equal chance of being picked as a potential absorber of an enemy attack, and once picked, the % chance to mitigate is used to determine whether it triggers.

[h3]Prompt context[/h3]
Added all equipped items to the story context, as well as the name and description of the current item being used or triggered as armor.

[h2]Mood/energy[/h2]

Revised mood/energy prompt to reduce false positives and reduce the chance of a mood debuff per turn (which was happening too often). The new system is also moddable via the new fields in the survival bars, increased_strs and decreased_strs.

[h2]New cheat codes[/h2]

• opopen 1 - As mentioned above, causes all open-ended instructions to be unchecked
  
• task adv - Completes the current quest task
  
• task obs - Obsoletes the current quest task causing the quest to be rewritten
  

[h2]Misc[/h2]

• Tweaked new item prompt to slightly improve accuracy
  
• Tweaked enemy charmed prompt to reduce false positives
  
• Tweaked task completion prompt to reduce false positives
  
• Added world background to respawn enemies prompt
  
• Added world background to merchant items prompt
  
• Added logic to parse quests more tolerantly and accept a common wrong format produced by smaller models, in which one task can span multiple lines
  
• Added mitigation for a rare character duplication bug
  
• Added mitigation for crafting an item with an empty name so at least the save won’t be broken
  
• For NovelAI users: Migrated to the new NovelAI API to prevent future breakage
  
• Fixed bug when pressing Esc after picking an item to use
  
• Reduced new item false positives by preventing equipped item names from showing up as new items
  
• Tweaked consumable logic to apply the consumable effect even if event checks determined that bar went down that turn (this way energy drinks should work in combat even if you seemed exhausted during that story turn)
  
• Added brighter warning messages to all non-standard options that perform worse than Free Cloud, due to a couple negative reviews left by people who did not bother to try Free Cloud first
  

Note: At the time of this announcement, this update is currently in the Beta channel and will most likely be pushed to the Regular channel within 24 hours.

[h2]Better story illustrations[/h2]
Instead of passing your story turn verbatim to the image generator, the text generator will generate a visual description of the scene to send to the image generator. Improves coherence of story turn illustrations.

[h2]Skill check overhaul[/h2]
The previous skills system was designed around weaker LLMs which couldn’t follow instructions. The idea was a “shotgun” approach mimicking the chaos of how the brain works in real life, where each action would grow 3 skills, and statistically one of them might match something you learned previously. This approach was unpopular because it was too chaotic, and a little outdated with the capability of smarter LLMs. The new skill system outputs only 1 skill per action. The LLM is fed the entire list of skills in the universe and told to pick a closest matching one or generate a new one if none match (up to a maximum of 100).

Additionally, whereas previously the conversation skills were hardcoded, they are now AI-determined, which means you can have different skills in different contexts. For example, asking for information in a black market may use the skill Streetwise.

[h2]Better combat biasing[/h2]
Improved “power” string mechanic, which previously was only included in some actions and wasn’t influential enough. It is also moddable now. This message indicates the balance of power based on your level difference when in combat with an enemy. There are 5 variants (much stronger, slightly stronger, equal, slightly weaker, much weaker). In particular, the new version makes instant defeats less likely when fighting an enemy close to your level.

As a result of these changes, enemy plot armor was also removed, because it was interfering with story coherence in some cases.

[h2]Better item descriptions for looted, foraged, harvested, crafted, and storage items[/h2]
All of these items are now generated with name and description in one go, and also get the context of their parent object or place description. Please note: mods which edited these prompt files may stop working. To fix this, remove the affected prompts from the modded folder to force it to use the default prompt.

[h2]Better and more flexible NPC context[/h2]
Added NPC descriptions to the prompt, which largely fixes the inconsistent gender issue. As well, now NPCs mentioned in the story get that story turn added to their interaction history (as opposed to only the target of an interaction).

Added a checkbox for whether to disable NPC description auto-update. If disabled, it uses the old mechanism of having a separate “history summary”. The reason I leave it enabled by default is often as you get to know a character, their original description becomes outdated and needs to be re-summarized.

[h2]Added button for respawning enemies[/h2]
Respawning enemies is now player-controlled via a button instead of automatic. Also rewrote the prompt for spawning enemies to include both name and description for better coherence.

[h2]Better control/modding support of “full prompt string”[/h2]
Under chatgpt-prompts, there is a new file called full_prompt_str.txt which specifies how your story context is built (the part of the prompt preceding your action sent to the story generator). Editing this file influences all prompts which use ${FULL_PROMPT_STR}. Please note the lack of newlines is intentional, to account for the case where the string doesn’t exist (if it does exist, the newlines will be inserted by the game).

[h2]Added the following cheat codes:[/h2]
npc [name]: Spawn an npc with the given name
ene [name]: Spawn an enemy with the given name
thing [name]: Spawn a thing with the given name

[h2]Improved “minor failure” from conversation (no more Sheepishly)[/h2]
Instead of making you “sheepishly” say the thing, a minor failure in conversation now suggests to the AI that the character’s reaction may not be entirely what you intended. Thus minor failures are now enabled by default.

[h2]Better mood debuff mechanic[/h2]
Added a field called chatgpt_zero_status_str which inserts itself into the prompt when the survival bars reach 0. Mods can use it. This is less annoying than having a random % chance to trigger a low-mood story overriding your action.

[h2]Configurable basic action roll check minimum[/h2]
Changed "disable critical failures" option to a slightly more flexible one where you can specify the minimum roll of talking or using a consumable on yourself. Options are critical failure, minor failure, or minor success; default is minor failure.

[h2]Farming check[/h2]
Added an AI check for farming. The farm UI is now only available for places that could conceivably be used for harvesting/processing.

[h2]Misc[/h2]

• Tweaked a large number of prompts (event checks and otherwise) to improve accuracy. Some mods may be outdated. The easiest way to see which prompts were modified is to go into the relevant "default" folder and sort by Date descending. Most (not all) touched files have been modified; some may have been modified but reverted.
  
• Started limiting token length for some texts other than just full prompt string, with generous limits that won't affect the vast majority of users. If you go over, a warning will be displayed in the story view, but the game will proceed normally.
  
• Added learned name to the list of event checks that can be player-confirmable.
  
• Increased description length and greatly reduced chance of ambiguous gender of generated NPCs.
  
• Added creative mode option to name the location/region you’re going into.
  
• Added AI-determined vendor/merchant logic.
  
• Added logic using AI to determine whether a character went with you if you changed locations.
  
• Separated defeated vs dead entities to avoid situation where horde of zombies is defeated if one zombie is killed.
  
• Fixed bug where initial images were being generated twice in a row.
  
• Improved responsiveness of the console.
  
• Fixed bugs in NSFW fallback in cloud text gen services. Please report issues if you run into any more.
  
• Equipment slot check now includes item description.
  
• Fixed a couple bugs related to mention of a new location.
  
• Fixed “(blocked)” or “(locked)” issue with place names/descriptions.
  
• Fixed bug in map UI where a stale location image would be displayed in location details.
  
• Fixed bug in map UI where location description wasn’t scrollable all the way to the bottom.
  
• Added a bracketed note when NPC dies even if it was dead according to event checks (in case it was ambiguous in the story). The case where it was dead according to event checks is already covered by the extra turn implemented in the previous update.
  
• Fixed bug where enemy sometimes didn’t get a turn in combat.
  
• Prevented new location from triggering if it matches existing thing name.
  
• Slightly increased image gen prefix text box size.
  
• Fixed bug when pressing Undo after marking a task as complete.
  
• AI decision icons should now remain clickable if an error occurred during processing.
  
• Regions are now highlighted if matching the filter string in map search.
  

[h2]Improved Coherence[/h2]

• Upgraded all cloud services (Free, Silver, and Sapphire) to better models.
  
• Many entity descriptions are now also taking the recent story text into account, which solves most cases of entity descriptions not matching the story.
  
• If an NPC or enemy died via injuries instead of story-detected death, a system turn will be triggered to describe their death, so that the AI can understand they're dead.
  
• Place descriptions will be automatically updated if they were mentioned again in the story.
  
• NPC descriptions will be automatically updated based on their interaction history (can be disabled by disabling interaction history in Options).
  
• Entering a location now causes all entities to be generated with their description (as opposed to generating the description later). This reduces the chance for entity descriptions to be out of place for the world.
  
• When you enter a new place via a story turn, the story turn will be taken into account when generating the location’s entities.
  
• Improved location change event check accuracy by including original location in prompt.
  

[h2]Improved Player Control[/h2]

• Multiple event checks are now player-configurable in terms of whether they require confirmation via a button press. These now include combat, task completion, task obsoletion, new item, new character, new follower, and new location.
  
• Quest V4 tasks can now be viewed and edited, both from the new game view as well as in-game via Essence of Transformation or Creative Mode.
  
• Starting quest tasks can also be saved in world presets to share with others.
  

[h2]Misc[/h2]

• Fixed annoying issue where story turn illustration sometimes wouldn’t go away even after mouse stopped hovering over it.
  
• Fixed bug where “computer room” ambience wouldn’t play.
  
• For autogen world preset, the last typed in text is remembered so you can easily retry your generations.
  
• Fixed writing style not showing up in prompt for Kobold API users.
  
• Fixed bug when pressing Undo while unblocking a blocking thing.
  
• Added prompt text to help AI coherence when unblocking a blocking thing.
  
• Fixed issue experienced by a small portion of local image gen users, where images wouldn’t show up in the game.
  

[h2]Quests V4[/h2]

Hard to believe we’ve already been through 3 iterations of quest systems. V1 was just a single task you could perform at a single location. V2 and V3 were multi-stage and a bit more immersive, but still designed for an “older” era (just 1-2 years ago) when LLMs weren’t nearly as powerful. Quests V4 is the next logical step to take this game to the next level of immersion, and utilizes the capabilities of modern LLMs for storytelling much more effectively than before.

When you start a new quest, an entire plot line is now generated using your world background and story so far. As you complete the tasks, you reveal more of the plot.

The plot is not set in stone. At any time, if you express interest to take a different path, or if the AI thinks a task became impossible (e.g. an essential NPC died), the remaining tasks will be regenerated.

The last task of each quest segment contains a choice, where each choice leads to a different ending.

At every turn, if you are tracking a quest, an event check is included to see if you progressed along the current task. Possible outcomes are: No change, instant completion, partial completion, or task became obsolete (due to player picking a different strategy, or task becoming impossible).

At each stage when you complete a task, an extra story turn may be generated which reveals more of the plot line and introduces you to the next task.

Any action you take has the potential to progress your current task, but if you feel stuck or uncreative, just click the orange button which will generate 3 options somewhat relevant to your current task.

Finishing a quest will create a persistent world status effect that lasts for a configurable number of turns (defaulting to 10).

Quests V4 applies only to new saves. Older saves can continue questing on the Quests V3 system.

A number of issues were uncovered in the Beta and since fixed, and the feature is now being rolled out to the Default build. The fixes include:

• Fixed story turns being broken
  
• Fixed tooltip text going off screen
  
• Changed format of task generation to reduce chance of error when using smaller models
  
• Increased token limit for local API options to reduce chance of error when generating quest tasks
  
• Tweaked prompt for story result from task result, reducing chance of a disconnect between finishing one task and starting the next one
  

Please report any remaining issues in the Discord or Steam discussions.

[h2]Event Checks UI Overhaul[/h2]

Event check messages have been converted to icons, which reduces clutter in the text view. Mousing over the icons provides more info about the triggered event check. Additionally, some event checks are now player-controlled and require you to click on them to confirm them, which helps mitigate false positives.

[h2]Story turns minor overhaul[/h2]

The logical representation of the text view has been improved, so now editing a story turn can finally cause it to be reflected in the text.

[h2]Misc[/h2]

• Generic exception message now includes a link to Player.log (finally) which you should include in any bug reports
  
• Removed deprecated “local” option. You may of course still play locally using one of the supported local text gen options (kobold or oobabooga).