Session settings

Advanced settings can be chosen before a session starts and changed later in Chat settings. They are stored in settings.json beside the session database and read again before use.

The Fast preset button at the top of the form fills in settings that trade depth for speed: no thinking for the helper agents, no scene preview, no verification pass, no memory look-ups by characters, at most two rounds of three characters, no off-stage rounds, no automatic reflection and a closed world. Nothing is saved until you save or start the session, and each setting can be changed afterwards.

Important groups include:

• Rounds: maximum rounds and characters activated for one player message.
• Memory: imperfect memory, recall count, lookup tools, retrieval breadth, keeper verification and how quickly a perception fades after a belief has digested it.
• Reflection: whether the characters sort what they have taken in into what they believe by themselves, every how many turns, whether a change of scene also starts it, the fewest turns between two reflections, how many characters one takes up and how much new a character needs.
• Decision model: who judges the statements that triggers wait on and who checks a belief against what a character took in since: the language model (the default), a decision model, both with the decision model where it is sure, or both while only the language model's answer is used, so that you can see how often they agree. Without a decision service the language model is used. Two more settings turn on checks of what the game master delivers, both off by default: that the narration does not speak, act or think for your character beyond what you wrote, and that the output does not step out of the story (a refusal, a warning, a disclaimer). A fault sends the narration back to the game master to rewrite, at most twice, after which it is delivered as it is. Two further checks need a decision model and have no language model version; both are off by default, and off the story plays as it always did: when the game master judges your action allowed, the decision model also checks it against every rule of the world and your own rules in one fast pass, and a rule it thinks is broken is put to the game master once to judge again; and the output is checked for stating hidden lore your character does not know (lore listed for other characters only), which sends it back like the other checks. Without a decision service neither check is made.
• Context: explicit token limit or a ratio of the selected model's context window, plus short- and long-term summary sizes.
• Authoring: whether the GM may ask the world author to create missing entities and how many may be added per turn.
• Images: generation mode, reference handling, reuse, interpretation and captioning models, and the hires fix (render again larger from the first result, when the image service can).
• Prompts: prompt language and overrides.
• Preset: the preset the story's narration instructions came from. In Chat settings a Preset section says whether the story still matches the library and can switch, re-apply or remove it. At the start the preset also fills the narrator and character temperatures you did not set.
• Interface pack: which interface pack the story uses, or none; the Scene pack that ships with the plugin can be named in any world. In Chat settings this is a section of its own that also reports problems and can bring in the world's newer packs.
• Reasoning and sampling: child-agent reasoning and role-specific model sampling.
• Dice: whether you see what the game master rolls: a collapsed card under the narration to open when you want to check, or nothing, to stay in the story. The dice are rolled either way. Start a message with /roll 2d6+1 pick the lock to roll yourself; the result is told to the game master as a fact.
• Tools: the optional tools each kind of agent (game master, characters, record keeper) may use, as switches. A choice applies to every agent of that kind, not to one character. In Chat settings it is the Tools section, saved as you switch; a world can set its own in the editor, and a story starts from what its world says. Besides the game master's dice tools, a character's look-ups (recall_my_memories, recall_my_known_entities, recall_my_lore) and four of the record keeper's (search_events, list_threads, list_relationships, list_items) can be switched off to make their requests cheaper; all are on by default. The record keeper's read_records and edit_records, and its place, faction, item, agenda and variable look-ups, are not optional. Names that this version does not know are kept in the file and not shown, so a world moved between versions with more or fewer tools still loads.
• Tools from other plugins: under each kind of agent, a folded list of the tools other plugins registered for everyone (file or shell access, an MCP server such as a search engine). They are off unless you switch them on, per kind of agent: the game master, the characters (all of them), the record keeper. Without a switch an agent does not see them, not even on its first request, so loading such a plugin into the profile does not give your story's agents file or shell access by itself. A tool acts with the permissions its plugin has, so switch on only what the agent should be able to do. A switched-on tool that is not registered now is kept and has no effect.
• Models: the model of each kind of helper (characters, background characters, record keeper, suggestion writer, world and plot author) as provider/model; empty uses the chat's own. Characters run side by side and write short answers, and the keeper runs in the background, so a smaller, faster model often does as well there.
• Provider options: settings described by the selected embedding or media provider.

The What the world put in front of the story list above it shows what triggers delivered lately: beats for the game master and cues for characters, each marked waiting, taken up (the game master wove the beat in, or a character acted on the cue) or lapsed (nobody did in time). Use it to see whether a trigger had an effect.

The Decision model section shows whether a decision service is available and which, and, for each site that can use one, what it is set to and how it has done in this story: how many judgements the decision model made, how many it handed to the language model because it was not sure, how many fell back because it could not be used, how often it agreed with the language model when both were asked, how long each took on average, and the latest disagreements. The counts stay with the story when it is taken back.

The Reflection section shows when the characters last sorted their memories, what each came to (new beliefs, changed, let go) and why a pass failed when it did. Reflect now starts a pass at once, also when automatic reflection is off; it runs in the background, and cannot be started while a turn is being played or another pass is running.

The Session doctor section opens a separate chat for this story (see Sessions): an assistant that reads what the story has recorded, looks for what does not fit, and proposes fixes that you approve one by one.

What the helpers did lately lists each helper's model, duration, tool calls, tokens and failure reason. Open a run to see its requests, prompt size, time to first output and whether the output limit cut it off.

Tokens per turn adds up the game master and every helper that worked for that turn, including background work completed after delivery. Cached input is shown separately when the provider reports it. A missing cache figure is labelled "cache not reported", not zero. DSH does not know the provider's prices, so this page does not estimate cost. Taking the story back removes usage for turns that no longer count.

The chat's Trajectory tab holds the game master's requests, timing, usage and failed tool calls. A refused call includes its reason there.

Where each turn went keeps diagnostics for the latest 30 turns. Open a turn to inspect:

• time spent waiting, generating and calling tools, split between work before and after delivery;
• the game master's largest prompt, divided into instructions, tools, scene context, current input, current steps and earlier conversation;
• refused or repeated calls, steering messages and outputs cut off at their limit;
• earlier events recalled for the player input, why they matched, and how many matches were left out;
• what each character received, compared with information the story did not give that character.

Prompt sizes are estimates from the text because the model service reports only the whole request. Turn on Keep what the helpers were told to include the names behind the character-access counts.

When you need to know why a helper said or wrote what it did, turn on Keep what the helpers were told (diagnostics.keepRuns), play, and open a run with Show: you see its instructions, what it was asked, each tool call with what came back, and the result it submitted. For a character it also lists what it was reminded of and why each memory came back: the cues (names, shared words, place), the scores of the word and meaning searches and of the reranker when they ran, how much it mattered, how long ago it was and how much of it is still in mind. The latest 60 runs are kept in the session's folder. The texts include what characters know in private, so leave it off when you are not looking for a problem.

Each run in the activity list shows its brief in parts (each paragraph that opens with a heading line, such as Where you are, with its estimated size, largest first), whether or not the chat keeps runs. A character acting while the player is elsewhere is marked off stage, and a background run is listed under the turn it began in, not the one it happened to end in. In the turn breakdown, each character's row also has a Show link to exactly what it was told for that activation when the chat keeps runs; without that setting the row says so.

Story statistics (Chat settings, read when you open it) counts the shape of the story from what is kept: turns on the way the story went, how long you waited for the game master, how much story time has passed, scenes (stretches of events in one place), who spoke how often and what share of the words, how much is narration, threads open and closed, relationships that changed in play, turns played more than once from the same state and turns set aside, and the tokens the game master and the helpers used. It judges nothing and is read-only.

An invalid external edit is logged and the last valid settings remain active. The complete defaults are listed in Configuration reference.

Recall check

Open Recall check from the right sidebar and describe a possible moment. Choose one character, or leave it on everyone with the player. The page shows the memories each character would receive, their clarity, and the factors behind their ranking. It also lists available background and counts of known people, places, things and routes.

Meaning search requires stored vectors and an embedding service; word search still works without them. A recall check does not play a turn, call a language model, or make the returned memories last longer.

Settings with no form field used to be edited only in settings.json; the form now has them all: the game master's and the helpers' output caps and the output room, whether helpers must answer with a tool call (sampling.tools), the sizes of the short-term and long-term summaries and how long a cue stays pending. A model that thinks without end stops at the game master's cap (default 12000 tokens) instead of running for minutes.

Import and authoring assistants

Importing a card, a preset or a lorebook, translating a world and the world author's own helpers run as small agents, and they are set once for the whole plugin, not per story: Settings → Role Play → Imports and authoring assistants. Changes take effect at once.

• Thinking level: how hard they think, as a level the model offers (for example off). A model that thinks without end never gives an answer, so set off when an import stalls. It also applies to the authoring chats themselves (the world, preset and lorebook assistants you talk to), whose model must offer the level. Empty uses the level of the story helpers below it, then the chat's own.
• Thinking level of story helpers: the level the characters, record keeper, authors and suggestion writer of a story get when the chat's own setting is empty.
• Model: a provider/model for these helpers; empty uses the chat's model.
• Temperature, most tokens per request and helpers at once: how freely they write (low by default, because they copy and sort), how long one answer may be, thinking included, and how many run side by side.
• Most tokens per request is a ceiling: each job also has its own, smaller one for the answer it needs, and a helper that writes text up to its limit instead of giving its result is stopped at once rather than reminded. An authoring chat's reply that runs out of room is steered back to act or to sum up.
• Minutes before a helper is stopped: a helper that has no result after this long (15 by default, 0 never) is stopped and the step says why, so a run-away helper cannot hold an import for hours. Stopping the chat's turn stops them too.
• Decision model for sorting a card: with decision, and a decision model selected in the providers, the pieces of a card that are only instructions or notes are settled by the decision model when it is sure (confidence, 0.7 by default) and are not put to the language model; anything it is unsure of, or any failure, goes to the language model as usual. The survey reports how many pieces it settled. The deep quality check that asks whether a character sheet is enough to play uses it the same way.
• Changes by the world authoring chat: Ask for each change (the default) shows every change to the world and waits; Make changes automatically makes them at once and keeps anything replaced or deleted in the world's .trash folder. See the authoring assistant.

The instructions the helpers are given are the import prompts in the data folder's prompts and follow the language the world's authoring chat was opened in, as does the label of each piece they are shown and the chat's title.

Upload and download limits sit in the section below it and apply at once: the largest upload of a card, preset or lorebook (100 MB by default), the largest catalogue download (256 MB) and how many days the files an update replaced stay in the trash folder (30; 0 keeps them).