interactive-story/docs/GUIDE.md

AI D&D — design notes

How the engine works and why it is built this way. The README covers what the project does and how to run it; this covers the reasoning behind the parts that had a real choice in them.

Part 1 is the AI layer, which is where most of the design effort went. Parts 2 and 3 are what makes it a service rather than a demo. Part 4 is the web plumbing, kept short.

---

Contents

• Part 0 — Orientation
• Part 1 — The AI layer
  • 1.1 The turn pipeline
  • 1.2 Context assembly is a budget problem
  • 1.3 World state: the AI proposes, Python referees
  • 1.4 Output length, by measurement
  • 1.5 The memory bank
  • 1.6 Streaming
  • 1.7 The scripting sandbox
  • 1.8 Why there is no agent framework
• Part 2 — Data and correctness
• Part 3 — Production concerns
  • 3.6 Counting visits
• Part 4 — The web plumbing, briefly
• Part 5 — Measured results and known limitations

---

Part 0 — Orientation

What the thing is

An AI Dungeon clone. You write a scenario, then play an open-ended text adventure where a language model narrates the world. You type "I open the door", the model writes what happens next, and it remembers what came before.

Four things make it more than a chat wrapper:

1. A context engine. The model has a limited input window. The app decides, every single turn, which pieces of the story get to be in the prompt and which get dropped.
2. A world-state engine. The scenario declares stats (hp, trust, day). The model proposes changes to them each turn; a Python engine decides what actually sticks.
3. A story tree. The story is not a list. Any turn can hold more than one take, and writing below one that isn't the live one starts a branch that borrows every turn above the fork rather than copying it.
4. A scripting sandbox. Real AI Dungeon JavaScript scripts import and run, inside an embedded QuickJS interpreter.

Runs locally against Ollama for free, or hosted against any OpenAI-compatible endpoint.

The stack, and what each part is doing

Piece	What it does here
FastAPI (Python)	The HTTP server. Every URL like /api/adventures/3/actions maps to a Python function. Also does the SSE streaming.
SQLAlchemy	The ORM. Adventure, Action, Memory are Python classes; SQLAlchemy turns them into tables and turns attribute access into SELECTs.
SQLite / Postgres	The database. SQLite is a single file on disk (local). Postgres is a server (hosted, on Neon). Same code talks to both.
React (JavaScript)	The UI. Describes what the screen should look like for a given state; when the state changes it re-renders.
Vite	The frontend build tool and dev server. Bundles React into plain JS the browser can load.
httpx	The HTTP client used to call the model endpoint.
tiktoken	Counts tokens, so the budgeting is arithmetic rather than a guess.
QuickJS	A small embeddable JavaScript engine, used as a sandbox for user scripts.

The whole thing is one process in production: FastAPI serves the API and the built React files from the same port.

The shape of one request

you tap "Do"
  → browser sends POST /api/adventures/3/actions   {type:"do", text:"open the door"}
  → FastAPI route: check ownership, rate limit, turn lock
  → assemble the prompt
  → POST to the model endpoint with stream=true
  → tokens come back one at a time
  → each token is forwarded to the browser as a Server-Sent Event
  → React appends it to the screen as it arrives
  → when the stream ends: parse the state block, referee it, save the action

---

Part 1 — The AI layer

1.1 The turn pipeline

Everything that happens between "player pressed a button" and "text is on screen". Source: backend/app/routers/adventures.py (_generate_turn).

player input
  → onInput script hook          (user JS may rewrite or block it)
  → store the player action
  → retrieve memories            (embed recent story, cosine-rank the bank)
  → build_context()              (the budget allocator)
  → onModelContext script hook   (user JS may rewrite the whole prompt)
  → snapshot the exact prompt    (for the Insights panel)
  → provider.generate()          (streamed, token by token)
  → onOutput script hook
  → extract the fenced state block, referee the delta, strip it from the prose
  → save the action, stamping on it the script + world state it leaves behind
  → fire-and-forget: summarize + embed in the background

Two design choices are visible in that list before any of the details.

The prompt is snapshotted, not reconstructed. Every AI action stores the exact text that was sent to the model. That's what powers the Insights panel: open any turn and see each context component, its token cost, and why it was included. It's also what makes prompt bugs findable. The cost is storage (~74 KB per turn), which turns into a real performance problem later. See 2.5.

Every node records the state it leaves behind. state_after and world_state_after are attached to the action once its hooks and its delta have run, so a node carries the stats and the RPG values as they stood when that turn finished. Rewinding to before a turn is then a read of the node in front of it, which is the same operation as switching to another branch: one mechanism, and the reason undo, retry, and branch switching all put the numbers back instead of only rewriting text. (These were *_before fields originally; a tree needs the after value, because a branch's tip is what a reader standing on it should see.)

---

1.2 Context assembly is a budget problem

Source: backend/app/context/builder.py.

The problem

The model can only read so much. Say the budget is 8,000 tokens. A 200-turn adventure has far more story than that. Something has to be dropped, and what gets dropped decides whether the story stays coherent.

The naive version

Send the last N turns. That breaks in two directions: N turns of short exchanges wastes the window, and N turns of long ones overflows it. It also throws away the things that matter most: the premise, the character sheet, the fact that you promised the innkeeper you'd return.

What this app does

Split the prompt into fixed sections and elastic ones.

Fixed (always included, whatever they cost):

Section	What it is
narrator	The system prompt: how to write.
world_state_guide	The stat legend: what each stat means, its range, its bands.
world_state	Current values of every stat, plus NPCs in scene.
world_state_rule	How to report changes.
ai_instructions	Per-adventure steering.
plot_essentials	AI Dungeon's "Memory": the premise.
story_summary	The auto-maintained running summary.
used_memories	Top-K retrievals from the memory bank.

Elastic (fit into what's left):

Section	Rule
world_lore	Story cards triggered by keywords in recent text. Capped at 40% of the remaining budget.
history	Story turns, newest first, until the budget runs out.

The algorithm is three lines of arithmetic:

reserved  = sum of every fixed section + author's note + length hint + reminder
available = max(256, context_token_budget - reserved)

Then cards spend up to available * 0.4, and history spends available - cards_used, filling backwards from the newest turn.

The details that are decisions

Cards are capped at 40% of the elastic budget. Story cards are triggered by keyword match, so a scene mentioning six named things could pull in six lore entries and leave no room for the story itself. The cap makes the failure mode "some lore is missing" instead of "the model has no idea what just happened". Cards that don't fit are still reported to Insights with included: false, so the UI can show the lore that got squeezed out.

History fills newest-first and stops. Oldest turns fall out. This is the right direction because the old material is not actually lost: it has been summarized into memories and the running summary, which are in the fixed section.

If even the single newest turn is over budget, it gets hard-truncated rather than dropped. A prompt with no story at all would produce nonsense; a prompt with the tail end of the last turn produces something.

The author's note is injected 3 actions from the end, not at the top. AUTHORS_NOTE_DEPTH = 3. Instructions placed near the end of a prompt have more influence on what comes next than instructions at the top, because of recency. The author's note is a steering control ("keep it tense"), so it goes where steering works.

The world-state reminder goes dead last. The full emit rule lives up in the system block, hundreds of tokens away from where the model starts writing. A one-line reminder occupies the final slot. Same recency logic, applied to the thing most likely to be forgotten.

Past AI turns get their state block re-attached. The state block is stripped from the text before it's stored, so a replayed history would show the model twenty of its own past turns that contain no state block. That teaches it, by imitation, to stop emitting one. So _history_text() reconstructs the block from the stored delta and re-appends it when building history. The model sees its own pattern and keeps following it.

The performance trap hiding in this

Building the context needs the newest ~6,000 tokens of story. The obvious implementation reads adventure.actions, which loads every row of the adventure, and then throws 90% of it away. At turn 200 that was 839 KB of database reads to use maybe 70 KB, and it grew every single turn.

backend/app/context/history.py fixes it by serving three shapes directly from SQL: a tail, a slice, and a count. window_covering() fetches the newest 32 actions, measures their real token count, and if that's short of the budget it projects how many more it needs from the average length just measured rather than blindly doubling:

average   = tokens / len(actions)
projected = int(budget / average * 1.15) + 8

Each round fetches only what it doesn't already hold, so no row is read twice. Result: the same turn costs 129 KB instead of 839 KB, and stops growing at around turn 50. The cost is bounded by the context budget instead of by the length of the story.

There's a second rule in that module worth naming: if the actions are already loaded in memory, slice them instead of querying. The scripting pipeline hands the whole history to user scripts (AI Dungeon's API requires it), so on a scripted adventure the rows are already there. Issuing a query beside them would mean paying twice.

---

1.3 World state: the AI proposes, Python referees

Source: backend/app/worldstate/engine.py, plan/12-phase-rpg-world-state.md.

The problem

You want an RPG layer: hit points, trust, quest progress. Who owns the numbers?

Three options, and why two lose

Option A: a deterministic dice engine. The player types "attack the goblin", the engine rolls, applies damage, and the model narrates the result. This is what a real RPG does. It loses here because the action space is unbounded: the player can type anything, and mapping arbitrary natural language onto a fixed rules system is a harder problem than the one being solved.

Option B: the model owns the numbers. Let it track hp in the prose and trust it. This fails immediately. Models are bad at arithmetic, worse at remembering a number across twenty turns, and completely unable to obey their own frequency rules. Tell one "only change this every 5 turns" and it will change it every turn.

Option C, chosen: the model proposes, the engine disposes. The model narrates and appends a JSON delta of what changed. Python validates and clamps it before anything is stored.

narration: "The blade catches your shoulder. Gwen shouts and drags you back."

```state
{"player.hp": -15, "npc.gwen.trust": 5, "milestones.escaped": true}
```

The engine then applies, in order:

Rule	What it stops
Path must exist in the schema	Hallucinated stats
Value must be the right type	"a lot" instead of -15
Cooldown	Changing a stat more often than the scenario allows
Counters can't decrease	The in-game day going backwards
max_delta_per_turn	Losing 90 hp to a stubbed toe
Clamp to min/max	Negative hp, trust above 100
Milestones are sticky, true only	Un-completing a quest
Flags are two-way booleans	(Deliberately unrestricted: that's what flags are for)

Everything it rejects is reported, not silently swallowed. The Insights panel shows applied, clamped, and rejected paths per turn, and the chip under each narration shows what actually changed.

The reliability mechanism: word bands

A stat can carry bands:

"hp": { "min": 0, "max": 100, "initial": 100,
        "bands": [[0,20,"very weak"],[20,40,"hurt"],[40,60,"minor damage"],
                  [60,90,"healthy"],[90,100,"full health"]] }

Two things use them. The live state block shows the current band label, hp 55/100 (minor damage), so the model reads a word, not just a number. And the stat guide shows the whole ladder once per turn, so the model can see the full scale it's reasoning across.

The point: models reason well over semantics and badly over arithmetic. "He's badly hurt, so a solid hit should take him to very weak" is a judgment a model can make. "55 minus 22 is 33" is one it will get wrong often enough to matter.

The failure philosophy

Nothing in the world-state engine raises. A malformed delta returns {} and the turn continues. The parser is deliberately tolerant: it strips trailing commas and leading + signs on numbers, both of which weaker free models emit and strict JSON rejects. It accepts a fence labeled state, one labeled json, or an unlabeled one, and falls back to a bare JSON object at the end of the text, but only if it parses into something that looks like a delta, so prose ending in } is never eaten.

This matters because the hosted demo runs on free-tier models. A stricter parser would mean a good model works and a free one doesn't.

One call, not two

The model narrates and emits the delta in a single request. The alternative, narrating and then making a second call to extract structured state, is more reliable per call and costs twice the latency and twice the rate-limit budget. On the free tier (20 requests/minute) that would halve the playable turn rate. The tolerant parser plus the terminal reminder was the cheaper way to buy the same reliability.

---

1.4 Output length, by measurement

The problem

max_output_tokens is a hard wall the endpoint enforces mid-sentence. Hit it and whatever is being written gets cut off. Since the state block is emitted last, the state block is what gets lost. The turn narrates fine and silently records nothing.

First attempt

Tell the model its budget: "keep this turn under about N words".

What the measurement showed

Average turn length went from 174 words to 246, and every run was longer than every unhinted run (n=5). Phrased as a budget, the number reads as a target to fill. The hint pushed turns toward the very wall it existed to protect.

The fix

Phrase it as a ceiling, and say explicitly that a typical turn is much shorter:

[Hard limit: this turn must not exceed 412 words. Write only as much as the
moment needs — a typical turn is much shorter. Finish the narration and append
the state block well inside the limit.]

Average came back to 170 words, and the state block survived at tight caps.

And the arithmetic around it

words = int((max_output_tokens - 50) * 0.75 * 0.90)

• - 50 (LENGTH_HEADROOM): tokens held back for the state block itself.
• * 0.75 (WORDS_PER_TOKEN): models can't count their own tokens, but they do follow a word budget. English prose is roughly 0.75 words per token.
• * 0.90 (LENGTH_BUFFER): a word budget is a suggestion the model overshoots, and the cap it protects is a hard wall. Aim 10% short so the overshoot lands in slack.
• Below 40 words the hint is dropped entirely: it stops earning its tokens.

---

1.5 The memory bank

Source: backend/app/memorybank.py.

The problem

Story history falls out of the context window as the adventure grows. Turn 4 said you promised the innkeeper you'd return. At turn 90 that's long gone from the prompt, but if you walk back into the inn, it should come back.

The three layers

raw turns   →  memories        →  story summary
(verbatim)     (every 6 turns)    (rewritten every 15 turns)
                    ↓
               embeddings  →  cosine similarity  →  top-K into the prompt

Layer	Cadence	Purpose
Memory	Every 6 actions, starting at 12	One or two past-tense sentences of concrete fact.
Story summary	Every 15 actions	A single ≤250-word overview of the whole plot, rewritten by folding in the new memories.
Retrieval	Every turn	Embed the last 4 actions (≤600 tokens), cosine-rank the bank, inject the top K (default 5).

Retrieval is the part that answers the innkeeper problem: the promise is a memory, the memory has a vector, walking into the inn produces a query vector near it, and it comes back into the prompt.

The decisions inside it

A memory attaches to the node whose block it ends on. Not to the adventure, and not to a position in a list of actions, but to a (branch_id, depth) coordinate. That is what makes "which memories described this turn?" an indexed lookup rather than a scan for rows whose covered range has fallen off the end of the story, and it is what makes memories inherit correctly across a fork: the ones above the fork point already sit on ancestors both lines read.

This is also how repair works. When a turn's text is replaced or removed (a retry, an undo, a deleted action), forget_node withdraws the memory attached to that coordinate and rewinds both marks to just before the stretch it covered, so the ground is summarized again from what the story now says. An earlier version instead held the newest action back a turn so it could never be summarized before it stopped being retryable. That is no longer needed, because the repair exists whether or not the invalidation happens at the tip.

Cursors only advance on success. Every AI call in this module is best-effort. If summarization fails, the function returns and the cursor is unchanged, so the same block is retried on a later turn. There is no retry loop, no dead-letter queue, and no backoff: the cadence is the retry mechanism. Failures are logged to the debug page.

Summarization is fire-and-forget, in a background task with its own DB session. The player's turn is already on screen; making them wait for a summarization call would add a second or two of latency to every sixth turn for no visible benefit. The task holds strong references to itself (the event loop only keeps weak ones, so a fire-and-forget task can otherwise be garbage-collected mid-run) and a per-adventure guard set stops two from overlapping.

Pinned memories count toward top_k. Pinned ones are always injected; unpinned ones fill up to top_k - len(pinned). Without that, 6 pinned memories plus top_k=5 injects 11 and blows the budget the whole context engine exists to respect.

A dimension mismatch scores 0.0, it doesn't crash. If the user changes their embedding model, old 768-dim vectors get compared against a new 1536-dim query. zip() would happily truncate and score garbage silently. An explicit length check returns 0.0 instead.

Eviction is LRU-ish, and evicted memories are kept. Over capacity (default 200), the least-used, least-recently-used unpinned memories are marked forgotten rather than deleted, so the UI can still show them and you can un-forget one.

Background calls never spend the shared demo key. The summarization and embedding providers are built directly from the user's own settings and never from the demo config, and their call sites are skipped when the turn is running on the demo key. Summarization is unmetered background spend; the demo key is server-funded. Both facts together would be a bill.

---

1.6 Streaming

The model produces tokens one at a time. Waiting for the whole reply before showing anything makes a 20-second generation feel broken.

Server-Sent Events (SSE) is the mechanism: an HTTP response that stays open and pushes data: {...} lines as they become available. It's one-directional (server → browser), which is exactly what's needed here. WebSockets would be a bidirectional connection for a unidirectional problem.

The chain:

model endpoint  --SSE-->  FastAPI  --SSE-->  browser  -->  React state  -->  screen

FastAPI reads the provider's stream, and for each chunk yields data: {"type":"chunk","text":"..."}. The frontend reads the response body with a ReadableStream reader, buffers on \n\n boundaries, and dispatches each parsed event.

Event types: player (the stored player action), reasoning (thinking-model traces, which stream into a separate collapsible panel with their own token budget), chunk (story text), stopped (a script blocked the turn), error, done.

Two production details that only show up when hosted:

• X-Accel-Buffering: no: nginx-style reverse proxies buffer responses by default, which turns a stream into one big delivery at the end. This header tells them to flush each event.
• The security-headers and body-size middlewares are written as pure ASGI rather than Starlette's BaseHTTPMiddleware, because the latter buffers the response body and would break streaming.

The empty-reply case is diagnosed, not reported as "empty". If a reasoning model streams thinking but no story text, it spent its whole budget thinking. The error says so and tells you which three settings to change.

---

1.7 The scripting sandbox

Source: backend/app/scripting/.

Real AI Dungeon scripts are JavaScript files defining modifier(text) and calling it as the last line, with globals like state, history, storyCards. To be compatible, this app runs the same contract in an embedded QuickJS interpreter.

The safety properties are mostly structural:

Property	How
No filesystem, network, or process access	QuickJS has none by default: nothing was removed, nothing was added
Memory cap	16 MB per run
CPU cap	2 seconds per run
No shared state between runs	A fresh Context per hook execution
A broken script can't break a turn	Every failure comes back as .error with text/state/cards unchanged; the pipeline logs it and continues

Data crosses the boundary as JSON: Python serializes {state, text, history, storyCards, info} in, and the script's results out. There is no object bridge to exploit.

One deliberate bug-compatibility: addStoryCard returns the new card's index, so the first card returns 0, which is falsy, so if (!addStoryCard(...)) misfires. That's upstream AI Dungeon's behavior. It's documented in the code and left alone, because matching real scripts is the whole point of the feature.

---

1.8 Why there is no agent framework

Graph-based agent frameworks (LangGraph and similar) earn their complexity with branching, cyclic, multi-step control flow: a graph of nodes where the path depends on what the model decides, with loops, retries, tool calls, and persisted state between steps.

This turn pipeline is a fixed linear sequence with exactly one model call. There is no routing decision, no tool selection, no loop. The "graph" is:

hook → retrieve → build → hook → call → hook → referee → store

Every turn takes that path. Adding a graph framework would mean carrying its state abstraction, its serialization model, and its debugging surface to express a straight line.

There is also a specific reason a framework's context handling wouldn't fit: the budgeting logic is the product. Buffer-window and summary-memory abstractions are opinionated about how to fit history into a window. Here the Insights panel exposes each context component, its token cost, and the trigger word that pulled it in, so the assembly has to be explicit and inspectable.

When it would be the right call: if the design went toward the two-call version, where narration is followed by a separate structured-extraction step, with a retry branch when extraction fails and a tool-calling path for dice, that is a graph, and hand-rolling it would get ugly fast.

This is not the same branching as the story tree (§2.2). "No branching" here describes control flow: the code path a single turn takes through the backend. That path never forks. It is not true of the data the app stores. When a player rewinds and plays a turn differently, that action creates a new branch in the saved history. One turn's execution is a straight line. The sequence of turns across a playthrough is a tree. The two claims are about different things and do not conflict.

---

Part 2 — Data and correctness

2.1 The domain model

User
 ├─ Scenario   (the template)      ── stat_schema, prompt, memory, author's note
 │    └─ StoryCard, Script
 └─ Adventure  (the playthrough)   ── world_state, script_state, head_branch_id/head_depth
      ├─ Branch  (one line of it)  ── parent_branch_id, fork_depth, lineage, name
      ├─ Action  (one node)        ── branch_id, depth, parent_id, live,
      │                               text, context_snapshot, state_after
      ├─ StoryCard  (its own copy)
      ├─ Memory     (text, embedding, branch_id, depth, use_count)
      └─ AdventureScript

The decision that shapes everything: template vs instance. A scenario declares what stats exist; an adventure holds what they are right now. Creating an adventure copies the scenario's story cards, scripts and plot fields into it, so editing a scenario later never mutates a game in progress. (There's an explicit opt-in "Update from scenario" flow for when you do want that, which diffs the two and shows you what would change.)

Same reasoning as instantiating a class: shared definition, independent state.

2.2 The story is a tree

The largest structural change the project has had, and the one with the most reasoning behind it.

The problem

The story used to be a list, and a mutable one. Retry rewrote the last entry in place; undo and delete removed entries from the middle. Everything derived from the story, such as the memories, the running summary, and the two marks saying how far each had got, was indexed by position in that list, and a position means something different after anything in front of it is deleted.

That single fact produced a family of bugs that all looked different:

• Deleting a middle action slid a never-summarized action down into the "already covered" range, so a recent action silently never became a memory.
• Discarding a memory left its actions behind the mark, describing nothing.
• Retry rewrote an action's text after the mark had passed it, so its memory described narration that was no longer in the story.
• The retried row was still attached to the adventure while its replacement was being written, so the model was shown the attempt it was meant to replace and wrote a continuation of it. That exclusion had to be threaded through four separate readers.
• Attempts lived in a JSON array on the row with a mirrored copy of the live one in the ordinary columns, which is a repeating group and a denormalisation in one.

Each was fixed where it was found. The pattern only becomes visible when you line them up: they are all the same bug, and it is that the story is a list nobody may reorder.

The shape

Make the story a tree, and none of them are reachable.

Every action is a node with a branch_id and a depth. A branch is one line through the tree; it holds the nodes played on it and borrows everything before its fork point from its ancestors. Nothing is ever copied, and apart from an explicit delete, nothing is ever removed.

branches(id, adventure_id, parent_branch_id, fork_depth, lineage, name)
actions(id, adventure_id, branch_id, depth, parent_id, live, text, …, state_after)
memories(…, branch_id, depth)
adventures(…, head_branch_id, head_depth)

depth is a position along a path, not a global turn number: A4 and B4 are two alternatives, not two turns. Reading branch C, whose tip is at depth 7 and which left B at 5, which left A at 3:

SELECT * FROM actions
WHERE (branch_id = 'C')
   OR (branch_id = 'B' AND depth <= 5)
   OR (branch_id = 'A' AND depth <= 3)
ORDER BY depth DESC LIMIT 32

→ A0 A1 A2 A3 B4 B5 C6 C7.

Why branch_id + depth rather than parent pointers alone. Parent pointers are the obvious way to store a tree and the wrong way to read one: reading a story would be N round trips up a chain, which throws away the windowed history work (§1.2) that made a turn's read cost flat. Depth replaces the old index as the ordering key, so the reads keep the shape they already had.

The lineage, and why fork count doesn't cost anything

The OR-clause above is not reconstructed per read. It is stored on the branch row as lineage, for example [(C, ∞), (B, 5), (A, 3)], computed once when the fork happens, from the parent's lineage plus one entry. context/lineage.py is the only module that knows how to turn it into a query, which is deliberate: one forgotten clause shows the wrong story and reports nothing.

Two properties of the shape do the real work:

• The ranges are disjoint and descending. A branch's own nodes always sit deeper than its fork point, and each ancestor is capped at the fork depth of the branch beneath it. So ordering the whole clause by depth DESC reads entry 0's nodes, then entry 1's, then entry 2's, which means a tail read can use the newest few entries and stop.
• Clause count is bounded by the context window, not by fork count. A 200-fork story whose newest branch is 40 turns long reads with one clause, because the window is covered before the second entry is reached.

A branch stores no story of its own, so a fork costs an id, a parent, a fork depth and a cached ancestry. Measured on a 40-turn story forked twenty times against the same story flat: a page load of 31,652 B against 31,433 B, a 1.007× ratio, or about 103 bytes per branch. No migration, no vacuum, no copy.

What a player actually does

None of the above is what the screen shows. In the player's words:

Any turn can gain another take. On an AI turn that means regenerate; on your own message it means type something else. Stepping between takes with ‹ 2/4 › is free: the story below simply empties, because that take has no children yet. A branch is created when you write below a take that is not the live one, never before.

That rule collapses two operations into one and deletes a distinction from the UI. The first version of this screen had a chip that switched at the tip and only previewed above it, with a second button to take that line: one control whose meaning depended on where the reader was standing. The rule above replaced it with a pager that only ever steps, a fork button on every turn, and no tip-versus-past distinction at all. The distinction survives in the implementation, where it decides whether a write needs a branch: at the tip the attempts are still leaves nobody has built on, so taking one is a switch and no branch is created.

Takes are grouped by parent, not by coordinate

The load-bearing detail, and the one that is not obvious.

The natural way to find "the other takes of this turn" is by coordinate: same branch, same depth. It is wrong in both directions:

B ── C   C1   C2              <- three takes, one parent (B)
          │    └── D1' D2'    <- two takes, parent C2
          └── D1 D2 D3        <- three takes, parent C1

Standing on the C2 path at that depth must read 2/2, not 5. Coordinate grouping gets that one right by accident, because writing under a non-live take forks and the two sets land on different branches. It gets C wrong: once C has been forked onto a branch of its own it is alone at its coordinate and reads 1/1, having lost C1 and C2 from a pager that must still say 1/3.

So a node carries parent_id, read for nothing but this. The alternative, making a branch's fork point a node rather than a depth so a promoted take never moves, was rejected: the whole point of lineage is that a read is an OR-clause per branch instead of a walk up parent pointers, and re-pointing the fork at a node changes path resolution itself, dragging in the cursors, memory depths and both bundle formats. parent_id is one indexed lookup, never a walk, and nothing about how a path resolves changes.

Cursors become anchors

The two marks, how far the memory bank has got and how far the summary has got, used to be counts. A count is a position in a list, and every rule about sliding them, rewinding them and translating between positions and Action.index existed to patch up the fact that the list moves.

A cursor is now an anchor: (branch_id, depth), the node up to and including which the work is done. Deleting an action does not move it, because a depth is a coordinate along a path rather than a slot in a list. "What is not covered yet" becomes a question about the story instead of about a list index, and it answers correctly whatever has been deleted in front of it. The branch half is what makes it survive forking: a depth alone is ambiguous once two branches both have a node 41.

position_of_index, note_action_removed, settled_story_actions and the cursor-rewind machinery were deleted, not left unused. So was the one-turn memory holdback that existed because a retry could rewrite an action the mark had already passed.

Derived work attaches to the node that produced it

Generalize the rule and a lot falls out: anything derived attaches to the node that produced it. A memory covering depths 37–42 attaches to that branch's node 42 and is invisible to any path that does not run through it. Shared ancestors are therefore shared automatically, so a fork needs nothing recreated: the memories above the fork point are already on the ancestors both lines read.

The subtle case is the memory sitting at the forked coordinate. The first cut moved it onto the new branch and re-anchored the marks naming it. Both are wrong for the same reason: that memory describes whichever attempt was live at that coordinate, which is the one staying on the parent. The right answer needs no code: the lineage caps the parent one depth short of the fork, so the memory is simply out of range from the new branch, invisible to both the retrieval clause and the anchor read. The new line summarizes that ground again, from the text it actually tells.

Hand-written memories obey the same rule. One used to carry a NULL depth, described as "belongs to the adventure rather than to a path". That sounds harmless and is not: a NULL is a coordinate no fork can cap, so a note typed on one line followed the reader onto branches whose events it never described. They are anchored at the head instead: the story you were reading when you wrote it.

Deleting, and why the branch UI was a hard dependency

Nothing is ever auto-pruned. That is the guarantee the whole design rests on, and it is also why branch management could not be a nice-to-have: without a way to delete a line, storage grows without limit.

The delete rule has two halves and the second is easy to miss. Refusing to delete the line being read is obvious. The other half is refusing any line it was forked from: parent_branch_id cascades, so deleting an ancestor takes the head with it and leaves head_branch_id pointing at a row that is gone. One membership test against the head's own lineage covers both, because a lineage already names itself and every branch it borrows from. The server is the authority; the client computes the same set only so a button can say so before it is pressed.

The migration, and what it deliberately did not do

There is no feature flag. A linear story is a tree with one branch, so the intermediate states were not half-migrated: they were the same product with a superset schema underneath, which made "existing adventures are unaffected" a literal, testable pass condition at every step. A flag would have bought two live code paths through the context builder, the memory bank, undo and retry at once.

The legacy columns (index, variants, variant_index, the two *_before snapshots) were kept unread for a release rather than dropped with the migration that stopped using them, so that a redeploy of the previous build is still a way out. Dropping columns is the one step that isn't.

One operational note that generalizes: on Postgres, a migration that rewrites every row of actions roughly doubles the table, and only VACUUM FULL gives it back: 79 MB reclaimed in 5.5 s on one occasion. But bloat scales with the heap, and context_snapshot is 94% of this table and lives out of line, so a migration touching only small columns reuses the existing TOAST pointer and costs a tenth of that. Read the sizes from sum(octet_length()) per column, not from n_live_tup, which is a stale estimate in exactly the direction that makes bloat look smaller.

What this is honest about

• The two marks are one pair on the adventure, not one per branch. Switching branches makes the mark on the line being left unreadable from the new one, and that ground is summarized again. It answers "nothing covered", which is the safe direction: redo the work, never skip it. But switching back and forth costs AI calls. Per-branch cursors are the fix if it ever matters.
• Story cards stay adventure-wide. A card invented on branch B shows on branch A. Event-sourcing card changes onto nodes was considered and rejected.
• Editing an already-summarized action still leaves its memory stale. The machinery to fix it now exists: an edit could write a sibling take and switch to it, which is a retry the player typed. It does not do that yet.

2.3 Undo and retry that actually rewind

Most implementations of undo delete the last message. That's wrong here, because a turn mutates three things: the text, the scripting scoreboard (script_state), and the RPG stats (world_state).

The mechanism: every node carries state_after and world_state_after, deep copies of what the adventure looked like once that turn had played. Rewinding to before a turn is a read of the node in front of it, so undo, retry and a branch switch are the same restore. The cooldown clock comes along for free: it lives inside the world state, in _meta.last_changed, so each line of the story carries its own without anything having to know there is one.

Nothing a retry replaces is discarded. The old attempt stays as another take of that turn, a sibling node at the same coordinate with live set to false, and the pager steps between them. Retry is not a special case: it is the tree, with the branch not yet created. See 2.2.

Three details that are easy to get wrong:

The turn being retried is excluded from its own context. Its takes are still attached to the adventure, so without exclude_action_id the model would be shown the attempt it is replacing as established story and would write a continuation of it. The exclusion had leaked into four readers, not one: history replay, story-card trigger matching, in-scene NPC detection, and the memory-bank similarity query. The invariant is worth stating flatly: anything reading the story during generation takes the exclusion.

A retry reuses the turn's depth, not the next one. Cooldowns are measured along the path, so allocating a new depth would advance the clock the cooldown rules run on and a retry would quietly unlock stats that should still be waiting.

delete_turn used to mean "every take at this coordinate". Once a take can be forked onto a branch of its own, the group spans branches, and undo reached across and deleted a take belonging to a line nobody asked about. Anything that reads a take group and then writes has to say whether it means the turn or the coordinate.

If the regeneration fails, the rollback is reversed. generate_turn wraps the generator in a try/finally: if it ends without saving, whether from a provider error, an empty reply, a script stop, or the browser hanging up, the previous take is put back in charge. Otherwise the state on the server would drift from the text still on the user's screen.

2.4 The turn lock

One turn at a time per adventure. Double-clicking "Continue" must not run two generations.

The subtlety: the check has to happen in the request phase, not when the SSE generator first runs. A StreamingResponse doesn't start iterating its generator until the response begins, so a check-inside-the-generator lets two rapid requests both pass before either one claims the slot. And because sync FastAPI endpoints run in a threadpool, the test-and-set needs a real threading.Lock.

def acquire_turn_lock(adventure_id):        # in the request handler
    with _active_turns_guard:
        if adventure_id in _active_turns:
            raise HTTPException(409, "A turn is already generating…")
        _active_turns.add(adventure_id)

async def with_turn_lock(adventure_id, gen):  # wraps the SSE generator
    try:
        async for event in gen: yield event
    finally:
        _active_turns.discard(adventure_id)

In-memory, so it's a single-process guarantee. That's honest for the deployment this targets: one Render web service. Two processes would need the lock in the database.

2.5 The 189x egress fix

The setup: Action.context_snapshot holds the entire assembled prompt for a turn, about 74 KB per row, 94% of the database.

The bug: every adventure load pulled that column for every action, to read two small fields out of it (the world-state delta, for the "what changed" chip, and the applied report). SQLAlchemy loads all columns by default.

The fix, in three parts:

1. Move the two small things that are needed for every action into their own column (Action.world_delta).
2. Mark the heavy columns deferred (context_snapshot, variants, reasoning), so they're only fetched when explicitly asked for.
3. Backfill the new column with dialect-specific server-side SQL, so the old data is extracted inside the database and never crosses the wire.

The result: one adventure load went from 38.5 MB to 0.20 MB.

The part that makes it stick: tests/test_egress.py hooks into SQLAlchemy's before_cursor_execute event, captures every statement the ORM sends, and fails if a bulk load ever names those columns again. The regression is caught by asserting on the SQL, not on a timing.

One more detail from that test's design: the count query is written as a real SELECT count(...) rather than query.count(), because SQLAlchemy's .count() wraps the entity select in a subquery, so the emitted SQL names every column, including the deferred ones. No bytes come back either way, but the database still has to read them, and a guard that greps SQL cannot tell the two apart.

There's a companion denormalization for the same reason: the pager has to know how many takes a turn has without fetching any of them, so variant_index and variant_count are cached on the row and refreshed by one function (attempts.renumber), precisely so they can't drift and the pager can't lie. variant_count is 0 rather than 1 for a turn nobody retried, because the question it answers is "is there anything to page through?"

2.6 Migrations, hand-rolled

No Alembic. An append-only list of (version, SQL) pairs, with the current version stored in SQLite's PRAGMA user_version or a one-row table on Postgres. 64 versions so far.

• A fresh database is created by Base.metadata.create_all() (always current) and stamped at the latest version; it never replays history.
• An existing database runs every migration above its stored version, in order.

Why this and not Alembic: for a single-file SQLite app that a user might have been running for months, the entire requirement is "add a column, don't lose their data". Alembic's autogenerate, branching, and down-migrations are machinery for a team with a staging environment. This is 250 lines and you can read all of it.

The constraint it creates is written at the top of the file: change models.py (so fresh databases are current) and append a pair here (so existing ones upgrade). Migrations 2–23 predate Postgres support and use SQLite-only syntax. This is harmless, because every Postgres database starts fresh and never replays them, but anything added since must run on both dialects.

One migration worth reading (#10, repairing duplicate action indexes) uses UPDATE … FROM with a window function rather than a correlated subquery, because SQLite may evaluate a correlated subquery against partially-updated rows and produce duplicates again while "repairing" them.

---

Part 3 — Production concerns

3.1 Two modes, one codebase

AIDND_MULTI_USER switches the whole app between two personalities:

	Local (default)	Hosted
Users	One auto-created "local user"	Guest on first visit, optional account
Auth	None: no cookies, no login UI	Signed session cookie
Rate limits	Off	On
Row caps	Off	On
API docs (/docs)	On	Off
Provider	Whatever Settings points at	User's key, or the shared demo key

The reasoning: a person running this on their own laptop should never be throttled by their own app, never see a login screen, and should get the interactive API docs. A hosted deployment needs all four of those to be the opposite. Rather than two builds, the differences are gated at each site.

Guests upgrade in place. A visitor gets a guest User row on first load. Registering sets email and password_hash on that same row, so every adventure they played as a guest survives with no re-parenting and no migration step. Three kinds of row share the users table: local (email NULL, not guest), guest (email NULL, guest), registered (email set).

Guests expire; accounts don't. One row per curious visitor adds up, so cleanup.py deletes guests idle for AIDND_GUEST_RETENTION_DAYS (default 5), measured as COALESCE(last_seen_at, created_at), because _touch only writes last_seen_at hourly and a guest minted by /auth/me has NULL until its second request. The filter requires both is_guest and email IS NULL, so upgrading in place is also how you opt out of expiry. It runs once at startup (the reliable trigger on a host that sleeps) and then every few hours.

It's a single Core DELETE, not db.delete(user): the ORM path would SELECT every adventure, action and memory into Python purely to delete them, and the FK graph is ON DELETE CASCADE from users all the way down, so the database can do the whole graph in one statement. Nothing a guest owns is visible to anyone else either: is_public is output-only, so shared content is exactly the seeded scenarios, which have user_id NULL and never match the filter.

3.2 The shared demo key

The demo lets people play with no signup and no API key, on a key the server pays for. That is a spending surface, so it's the most defended code in the project.

resolve_provider_config() is the single place the BYOK-vs-demo decision is made, and on the demo branch it pins two things:

• The model, pinned to a whitelist. A caller-supplied override or a hand-edited settings row can't aim a server-funded key at an expensive model. Anything unrecognized falls back to the first whitelisted model.
• The endpoint, pinned to the configured demo URL. Otherwise the key could be redirected to a URL the user controls and harvested.

Plus a daily per-user turn cap (default 20), checked before the player's input is stored so a capped player doesn't get their message saved with no reply, and counted only after a successful turn.

There's a defensive __post_init__ on the config object that raises if a demo config somehow carries a non-whitelisted model. The comment on it records a real bug: the check tests using_demo, not api_key == DEMO_API_KEY. Keying on the key value looks stricter but is wrong: the demo key is an ordinary OpenRouter key, so a user can legitimately paste that same key into their own settings as BYOK, and then every resolution raised, 500ing even GET /auth/me and taking the whole SPA down. using_demo is what actually means "the server is paying".

Background work (summarization, embeddings) is excluded from the demo key entirely: those are unmetered calls, and unmetered calls on a server-funded key is a bill.

3.3 Secrets

Everything derives from one server-side secret (AIDND_SECRET_KEY).

Thing	Mechanism
Passwords	hashlib.scrypt, N=2^14, r=8, p=1, per-password salt, constant-time compare. Stdlib, so no extra dependency.
Sessions	v1.<user_id>.<HMAC-SHA256>, no expiry: long-lived guest sessions are the point. A cookie can outlive a swept guest row; that resolves to a 401, which the frontend already turns into a fresh session.
Stored LLM API keys	Fernet (AES) encryption at rest, key derived from the secret, enc: prefix so legacy plaintext rows are recognizable and migratable.

The secret auto-generates into a file next to the database for local installs (zero config), but multi-user mode refuses to start without the env var, with an error message that explains why and gives you the command to generate one. Hosted filesystems are ephemeral; a regenerated secret on every deploy would silently log out every user and orphan their stored API keys.

A rotated secret makes stored keys undecryptable. decrypt_secret treats that as "unset" rather than raising, so the user just re-enters their key instead of hitting a 500.

3.4 Abuse guards

Guard	Value
Turn generation	10 / minute
Auth attempts	10 / 5 min, per IP
Guest creation	30 / 5 min, per IP (each guest is a DB row)
Script test runs	30 / minute (each costs up to 2s CPU)
Connection test	10 / minute (outbound HTTP to a user-supplied URL)
Adventures / scenarios / scripts per user	100 / 200 / 200
Actions per adventure	5,000
Request body	2 MB, 20 MB on import endpoints

Rate limits are keyed per user when one is known (accounts survive IP changes) and per IP otherwise, in fixed windows held in memory, with a pruning pass so the per-IP dict can't grow without bound. Import endpoints check bundle list lengths against the same caps live creation enforces, otherwise the cap is trivially bypassed by uploading a file.

Security headers on every response: nosniff, X-Frame-Options: DENY, Referrer-Policy: same-origin, and a CSP allowing exactly what the SPA uses: same-origin everything, inline styles (React needs them), and Google Fonts.

3.5 Deployment

One Docker web service on Render, serving the SPA and the API same-origin, with Postgres on Neon.

The Postgres decision was forced: Render's free tier has no persistent disk, so a SQLite file wouldn't survive a deploy. The database lives off-box on Neon's free tier.

Two things worth knowing about the free tier:

• The service sleeps after ~15 minutes idle, and the first request then takes 30–60s.
• /api/health deliberately doesn't touch the database, so a keep-warm pinger wakes the web service without waking the database. Waking a database around the clock costs far more than the cold start is worth.

CI runs the backend tests, the frontend lint and build, and a Docker image build on every push.

3.6 Counting visits

A hosted demo raises a question a local app never does: is anyone using it, and do they get anywhere? The answer is an owner-only dashboard at /analytics, gated on AIDND_ANALYTICS_EMAILS, a list kept separate from AIDND_POWER_USERS, since an unmetered tester is not automatically someone who should see the traffic.

Why it isn't a third-party script. The CSP allows script-src 'self', so a tracker would mean loosening it. Ad blockers block the popular ones, which silently biases exactly the technical audience this project is shown to. And none of them can see the measurement that matters here: a turn. The interesting funnel step is not a pageview.

Egress is the budget. After the 189x fix (§2.5) it would be perverse to add a feature that reads rows per request. So counts accumulate in a process-local dict and flush every 60 seconds as UPSERTs: a visit is a write and never a read. Storage is a generic (day, metric, label) -> hits counter plus one row per visitor per day for the funnel flags. Every dashboard query is a GROUP BY that returns tens of rows regardless of the traffic behind it, so a month costs a few kilobytes to read back. The cost of the buffer is that a hard restart can lose up to a minute; the flusher also runs on shutdown, and on a tier that sleeps when idle, the buffer it sleeps on is empty anyway.

The numbers are the server's, not the browser's. The client reports one fact, which page was viewed, and even that is normalized to a route (/play/12 → /play/:id) against a whitelist, so the page list cannot be polluted by anything a stranger posts. Everything that means something, such as a turn, an adventure, or a sign-up, is recorded by the code that performs it. That also fixes a blind spot: a failed turn is an HTTP 200 with a bad ending, so a status-code tally cannot see it, and a demo whose model has started refusing looks perfectly healthy from outside. turn_error is counted where the SSE error is written.

The funnel counts people, not clicks: a player who starts six adventures is one person who started an adventure, which is the entire reason the per-visitor-day table exists.

One smaller decision worth naming: error buckets are labeled by the matched route template, never the requested path. That gives one bucket per endpoint instead of one per adventure id. The reason it isn't merely tidier is that an unmatched path is entirely attacker-chosen, so labeling by it would let anyone mint rows.

---

Part 4 — The web plumbing, briefly

For the parts that are just how the web works, not decisions.

Frontend and backend are two programs. In development they're two servers: Vite on 5173 serving React, and FastAPI on 8000 serving the API. Vite proxies /api to FastAPI so the browser thinks it's all one origin, which avoids CORS entirely. In production there's one server: FastAPI serves the built React files as static assets from the same port.

SPA routing. React Router handles URLs like /play/3 in the browser without a round trip. But if you reload that URL, the browser asks the server for /play/3, which isn't a file. So SPAStaticFiles catches the 404 and returns index.html, letting React take over and read the URL itself. API routes are matched before the static mount, so they're unaffected.

Sessions. A cookie is a small value the browser stores and automatically attaches to every request to that site. Here it holds v1.<user_id>.<signature>. The server doesn't store sessions anywhere; it re-verifies the signature on each request, which is why there's no session table.

The 401 retry. If the cookie is missing or stale, any API call returns 401. The frontend catches that once, calls /api/auth/me (which mints a fresh guest session), and retries the original request. So a returning visitor with an expired cookie never sees an error.

React, in one paragraph. A component is a function that returns a description of some UI. useState holds a value; changing it re-renders the component. The streaming turn is the clearest example: each SSE chunk appends to a state string, React re-renders, and the text appears to type itself.

---

Part 5 — Measured results and known limitations

Measured results

Database egress per adventure load	38.5 MB → 0.20 MB (~189x)
Prompt snapshot size	~74 KB/turn, 94% of the database
Turn read cost at turn 200	839 KB → 129 KB, flat after ~turn 50
Cost of a branch	~103 B; 20 forks load at 1.007× the same story flat
Length-hint phrasing	174 → 246 words phrased as a budget; 170 phrased as a ceiling (n=5)
Backend tests	440, LLM mocked, real QuickJS engine
Schema versions	64
Sandbox limits	16 MB, 2 s CPU, fresh context per run
Context defaults	author's note at depth 3, cards capped at 40% of elastic budget
Memory cadence	memory / 6 turns, summary / 15 turns, top-5 retrieval

Two of the tests encode a performance property rather than a behavior: test_egress.py asserts on the SQL the ORM emits, and test_history_window.py asserts that the read cost stops growing with story length.

Known limitations

Deliberate trades for a single-user-first app that also happens to be hosted, listed so nobody has to discover them the hard way.

• Single process. The turn lock, the rate limiter and the summarization task all assume one worker. A second worker would need the lock in the database (a row-level advisory lock) and the rate limiter in Redis.
• No vector index. Retrieval does cosine similarity in Python over the whole bank. Fine at the 200-memory cap; at 10,000 it would want pgvector.
• Prompt snapshots are heavy even after the egress fix; they're deferred, not smaller. Compressing them or expiring old ones is the real fix.
• In-memory rate-limit windows reset on restart, so a restart grants a brief extra allowance.
• Background summarization is a fire-and-forget asyncio task, so it does not survive a restart. At real load it belongs in a queue.
• The demo key depends on a free-tier provider's daily cap, which the app can only detect after the fact by string-matching the 429 body.
• The two memory marks are one pair on the adventure, not one per branch. Switching lines makes the mark on the line being left unreadable from the new one, so that ground is summarized again. It fails in the safe direction: redo, never skip. But switching back and forth costs AI calls. Per-branch cursors are the fix if it matters.
• Story cards are adventure-wide, so a card invented on one branch shows on all of them.
• Editing an already-summarized turn leaves its memory stale. Replacing a turn withdraws what was derived from it; editing one in place does not.

Cleanup backlog

docs/self-review.md carries an open list of non-bugs (reuse, simplification, and efficiency items) kept deliberately separate from the correctness list, which is empty. The largest ones:

• Section.tokens is uncached, so the context gets tokenized two or three times a turn.
• onModelContext flattens system and story into one string before handing it to user scripts; if a script modifies it, the structure is gone and everything ships as user content. Passing structure through the hook would be better but would break AI Dungeon compatibility, which is the point of the feature.
• The import endpoints hand-coerce raw dicts instead of using Pydantic bundle schemas.
• The legacy pre-tree columns (index, variants, variant_index, and the two *_before snapshots) are still on actions, unread, kept for one release so redeploying the previous build remains a way out. Dropping them is a migration that rewrites every row, so it owes a VACUUM FULL actions; after it.

---

Source: github.com/parththakkar106/AI-DnD · Project page