When the LLM returns ontology attributes as plain strings instead of
dicts, set_ontology() crashes with "TypeError: string indices must be
integers, not 'str'" at attr_def["name"].
Two-layer fix:
1. ontology_generator.py: normalize string attrs to {"name", "type",
"description"} dicts during validation, so downstream code always
receives well-formed structures.
2. graph_builder.py: add isinstance guard as a safety net in
set_ontology() for both entity and edge attribute loops.
Co-Authored-By: Octopus <liyuan851277048@icloud.com>
When the LLM generates a <tool_call> block followed by a self-generated
<tool_result> block in the same response, the fake result must be stripped
before appending to message history. The real tool result will be injected
separately by the system.
This fixes a React Hallucination bug where models could fabricate tool
results that didn't actually come from tool invocations.
The API retrieval layer hardcoded reddit as the default platform in 11+ locations. When a Twitter-only simulation was run, all data retrieval APIs silently returned empty results because they looked for reddit_simulation.db which did not exist. This commit reads the simulation enable_twitter/enable_reddit config to determine the correct default platform.
Fixes#150
fetch_all_nodes already had a max_items guard (default 2000) but
fetch_all_edges had no such safeguard, allowing unbounded memory growth
on graphs with large numbers of edges.
Add _MAX_EDGES = 5000 constant and mirror the same loop-guard pattern
from fetch_all_nodes: cap the result list, emit a warning log, and
break pagination once the limit is reached.
Upgrade the transitive Python dependency NLTK from 3.9.2 to 3.10.0 to address CVE-2025-14009. The contributor title said 3.9.3, but the reviewed lockfile resolves 3.10.0.
- Add return type annotation (list[str]) to Config.validate()
- Add type annotations (msg: str, -> None) to logger convenience functions
- Add FileParser.is_supported() classmethod for checking file format support
- New Vue views: GodModeView (inject event, modify emotions with
preloaded current values, kill character with typed-name
confirmation) and WorldBuilderView (rules editor, locations with
cinematic atmosphere field, event log viewer)
- Shared sim-nav strip across Story/GodMode/World views
- Location schema extended with optional atmosphere field — a short
mood phrase that anchors the opening visual of every scene set there
- Translator's _format_world_locations surfaces atmosphere to the LLM
prompt so the field actually influences generation
- Routes added for /godmode/:simId and /world/:simId
Tests: 39/39 passing (added test_location_atmosphere_surfaces_in_prompt).
Frontend builds clean.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Six new endpoints under /api/narrative/*:
- GET /world/<sim_id> — full world state
- POST /world/<sim_id>/rules — replace rules list
- POST /world/<sim_id>/locations — upsert location
- POST /godmode/<sim_id>/inject-event — inject world event
- POST /godmode/<sim_id>/modify-emotion — overwrite emotions
- POST /godmode/<sim_id>/kill — kill character
Validation:
- 400 on missing required fields (description, character_id, rules)
- 400 on invalid round (must be non-negative int or null)
- 404 on character not found (from ValueError in handlers)
Smoke-tested via Flask test client — all 11 response codes correct.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When Marcus (the only round-2 actor) was killed, generate_prose
short-circuited on empty actions without calling the LLM, so
mock_llm.call_args_list only had one entry. Added Elena as a
second round-2 actor to exercise the filter with a live participant.
- PROSE_PROMPT_TEMPLATE extended with world_rules, world_events,
world_locations, and event_enforcement substitution fields
- EVENT_ENFORCEMENT_STRENGTH set to "hard" — opening line MUST
reference the most recent world event (aligns with existing
punchy, cinematic voice)
- Character summaries now resolve location id → name for prompts
- translate_round loads world_state and filters dead characters
from both the roster and the action list before prose generation
- User-supplied strings (rules, events, locations) brace-escaped
before str.format() to prevent KeyError on stray { or }
Tests: 36/36 passing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Full-pipeline test: fake simulation dir → actions.jsonl → two
translate_round calls → assertions on beat persistence, character
evolution, and translator offset advancement.
Final test suite: 20/20 passing across action mapper, character
engine, story store, translator (read + prose + orchestration), and
E2E.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Five endpoints under /api/narrative/*:
- GET /story/<sim_id> — full story so far
- GET /story/<sim_id>/round/<num> — single round
- POST /translate — translate a round on demand
- GET /characters/<sim_id> — roster with emotional state
- POST /characters/<sim_id>/init — bootstrap from OASIS profiles
Blueprint registered alongside existing graph/simulation/report
blueprints following the established pattern. Smoke-tested via Flask
test client.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
translate_round() ties together:
- reading the round's actions from actions.jsonl
- generating prose via the LLM
- updating per-character emotional state
- persisting the beat, characters, and file offset
Full test suite now at 19/19 passing across action mapper, character
engine, story store, and translator.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Module-level call_llm() wrapper (patchable by tests) drives
PROSE_PROMPT_TEMPLATE rendering with character summaries, actions,
previous beats, and user-selected tone. Prompt tuned for punchy,
cinematic, dialogue-heavy voice.
Tests: 5/5 passing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Reads one round's worth of agent actions from the OASIS log file,
tracking file offset so callers can resume across translation calls.
Handles missing files, malformed JSON lines, and non-target rounds
gracefully.
Tests: 3/3 passing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Character profiles now track a six-dimension emotional state (anger,
fear, joy, sadness, trust, surprise) that evolves based on actions
taken. Initial deltas for all 13 OASIS actions chosen as a balanced
baseline (see module docstring for tuning guidance).
Tests: 5/5 passing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Three JSON files per simulation under narrative/ subdir:
- story_beats.json: chronological story passages
- translator_state.json: per-platform file offsets
- characters.json: extended character profiles
Tests: 4/4 passing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New narrative package with action_mapper module that maps the 13 OASIS
action types (CREATE_POST, LIKE_POST, REPOST, etc.) to narrative verbs
and interpretations used when generating story prose.
Tests: 4/4 passing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
DEBUG defaults to True (FLASK_DEBUG env var is 'True' unless overridden) and the server binds to 0.0.0.0. When debug mode is active, Flask enables the Werkzeug interactive debugger, which allows arbitrary Python code execution via the browser-accessible debugger console if an exception is triggered. Combined with the public bind address this creates a remote code execution vector on any deployment that forgets to set FLASK_DEBUG=False.
Affected files: config.py
Signed-off-by: ChinhLee <76194645+chinhkrb113@users.noreply.github.com>
anadoris007
666ghj
octo-patch
ygh1254
IagoMelo01
MiroFish Contributor
karesansui
harshitanand
orbisai0security
lllopic
Sami Rusani
ChinhLee