Update V2 plan: event-driven targeting, local embeddings, unified tag system - #7
Open
nat-bishop wants to merge 22 commits into
Open
Update V2 plan: event-driven targeting, local embeddings, unified tag system#7nat-bishop wants to merge 22 commits into
nat-bishop wants to merge 22 commits into
Conversation
… system Major design updates to object-system-v2.md: - Replace polling-based target scanning with event-driven matching (runs only on object spawn or tag change, not per-frame) - Switch from cloud embeddings (Google text-embedding-004) to local model (all-MiniLM-L6-v2 via ONNX, <1ms, no PyTorch) - Unify tag handling: all tags get TAG_PHRASES, no static/dynamic distinction. Any tag can change at any time (Medusa petrification) - Identity = name + visual_description + tag phrases for all current tags - Behavior priority = list order (LLM decides character through ordering) - Minimal implied behaviors (only default wander for living + no behaviors) - HOSTILE stays LLM-visible for non-seeking hostile objects - Remove old design-object-system-v2.md (superseded) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…tion stages Replace all-MiniLM-L6-v2 (384-dim ONNX) with qwen3-embedding:0.6b (1024-dim via Ollama) for better semantic similarity. Flatten Stage 1/Stage 2 into Backend/Godot sections, remove backward-compat shims (V1 behavior conversion), add SIZE_PHRASES to embed strings, refine TAG_PHRASES formatting and examples. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…viors Replace the V1 trigger/action/target behavior system with a substance/element-based model. LLM now picks substance (16), optional element (7), 0-3 tags, and 0-4 verb behaviors. Traits derived from substance+element automatically. Semantic targeting via Ollama qwen3-embedding. Includes cleanup: persistent httpx client, re-embed debounce, model_dump for substance_profile, enum-derived verb values, dead code removal. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…te passthrough - Rename LLM_VISIBLE_TAGS_V2 → LLM_VISIBLE_TAGS and all derived constants - Strip "V2:" prefix from comments across Python and GDScript (20+ files) - Fix _reapply_physics() no-op: runtime tag changes (HEAVY, LIGHT, FLYING, BOUNCY, STICKY) now actually update physics via apply_tag_modifiers() - Eliminate PhysicsComposer.compose_physics_v2() one-line passthrough; call SubstancePhysics.resolve_physics() directly - Clean up SubstanceProfile sentinel defaults (-1.0 → 0.0) and remove dead sentinel checks on client side Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds a suite of embedding/targeting eval scripts capturing the investigation findings (embeddings can't reliably distinguish state/tag shifts). Updates embedding_eval.py thresholds and fixes a known-limitation test case expectation. Adds `make ollama` target to pull qwen3-embedding and verify Ollama is running. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Reflects new compiled_behaviors format, tag-filter targeting, stable identity embeddings, and asymmetric query embedding approach. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…embeddings Replace dynamic re-embedding with compiled behaviors that combine tag-filter mode (require_any_tags → pure boolean check) and identity mode (embedding dot-product). Identity embeddings are now stable (substance-based, no tags), and behavior targets use instructed-retrieval prefix for asymmetric matching. - Add target_tags field to BehaviorVerbSpec with ModelRetry validation - Compile behaviors at creation into self-contained dicts (verb, target, target_embedding, require_any_tags, exclude_tags) - Replace behavior_embeddings with compiled_behaviors on GameObjectBlueprint - Remove /api/embed endpoint and client-side re-embedding on tag change - Rename object_embedding_changed signal to object_tags_changed - Update VerbBehaviorSystem for dual-mode targeting (tag filter vs embedding) - Rewrite embedding evals for stable identity + instructed retrieval - Add targeting integration test (evals/targeting_test.py) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…t_tag Three-layer architecture: material rules (56 passive physics rules), behaviors + categories (creature intent via boolean matching), and planned spawn-time LLM enrichment (Phase 4, not yet built). - Add Category enum (CREATURE, FOOD, WEAPON, PLANT, VALUABLE, VEHICLE, TOOL) - BehaviorVerbSpec: target_tags → single target_tag from BEHAVIOR_TARGETS - CREATURE auto-derives LIVING, destroy verb auto-derives HOSTILE - LLM_VISIBLE_TAGS: 12 → 10 (LIVING/HOSTILE now derived-only) - Remove embeddings.py, Ollama dependency, identity_embedding - Remove 3 HOSTILE combat rules (VBS handles creature combat) - VBS: collapse dual-mode to single _matches_target() boolean - Player joins game_objects group with CREATURE category - Remove HOSTILE from field_config FIELD_TAGS Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds a CanvasLayer-based post-processing system for visual style iteration. Five fullscreen shaders (palette limiter, ordered dithering, sobel outline, crosshatch, pencil) can be toggled (F8) and cycled (F9) at runtime. Debug mode (F3) shows the active shader name. Materials are pre-cached to avoid allocation churn on toggle/cycle. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove obsolete embedding-based evals and targeting tests (replaced by category system). Add gate dataset, category accuracy metric, and benchmarking scripts. Update eval runner with --gate and --compare flags. Clear stale pre-V2 baseline. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Split design into focused docs: architecture (two interaction systems, pairwise cache, state interaction table), engine primitives (physics, 7 states with mechanics, affordances, verbs, destruction results), creation pipeline (call batching, budgets), and future features (compound sprites, object evolution). Includes archived V1 doc and inference benchmarks. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…flow - Rewrite LLM judge rubrics from 5-point generic to 3-point anchored scales with game-specific examples (Faithfulness, Delight, Visual Clarity) - Move 90 eval cases from Python to YAML (dataset.yaml) with schema - Add baseline management: --gate, --compare, --promote, --list CLI flags - Add gate dataset (14 cases, deterministic only, no LLM judges) - Fix judge truncation bug: remove max_tokens=300 that caused ~30-89% failures - Increase max_concurrency from 8 to 64 (SGLang batching, no rate limits) - Fix valid_behaviors checking wrong field name (target -> target_description) - Delete 7 dead files (research scripts, redundant runners) - Rename living_movement.py -> has_signature.py to match metric name Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…mplifications Major design changes from review: - Replace fixed behavior verbs with per-pair creature evaluation (staged: movement → action → parameters). Firefighter extinguishes fire, rescues people, chases arsonists — all from one creature. - Drop state interaction table (pairwise cache handles state changes with lazy invalidation) - Drop destruction results from core (death animation → remove, kept in future-features.md) - Cut states from 12 to 7 (removed unspecified: blessed, cursed, weakened, strengthened, petrified) - Add mechanics for each remaining state (DOT, armor reduction, etc.) - Rename local-llm-rules.md → llm-design-patterns.md with decomposition patterns (fan-out, sequential seeding, fan-out/filter/fan-in) - Add cache design, migration plan, future optimizations docs - Add benchmark scripts (outputs gitignored) - Delete archive directory Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add screened pairwise evaluation — single LLM call triages which object pairs are worth fully evaluating, cutting ~70% of N² calls. Add scene context to creature behavior prompts so creatures make contextually-aware decisions (firefighter cares about water when fire exists). Mark creature behavior cache as scene-local since results depend on scene composition. Add utility scoring priority resolution for competing creature behaviors — discrete target selection with perception-range-normalized distance falloff, same pattern as The Sims. Add batch screening as pattern 5 in LLM design patterns doc. Add triadic reasoning note to future features. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two plan variants covering the 10-phase dependency graph from V2 extraction through LLM question runner, pairwise interaction cache, creature behavior, Godot client rewrite, and eval system rebuild. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ates Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace V2's hand-authored substances (16), tags (31), elements (7), categories (7), behavior verbs (6), and 56 rules with: Backend: - Parallel 1-token LLM questions for physics (mass, friction, bounce, etc.) - Pairwise interaction cache (LLM-evaluated at spawn, looked up at runtime) - 7-state system (on_fire, frozen, wet, electrified, poisoned, corroded, molten) - Creature behavior via 3-stage LLM evaluation - CacheManager with SQLite backend for creation/pairwise/behavior caches - New creation pipeline, pairwise evaluator, and creature behavior engine Godot Client: - StateManager replaces TagManager with contradiction-aware state mutations - CollisionResolver uses pairwise cache lookup instead of rule matching - EffectExecutor rewritten as static class for pairwise result execution - Physics-derived visual animations (bobbing, breathing, squash/stretch) - State-driven visual effects (particles, shaders, color shifts, lights) - AffordanceSystem for carry/mount/climb interactions - FieldSystem for area-of-effect interactions - V3 blueprint format for all level seed objects and environment bodies Cleanup: - Delete V2 backend modules (substances, tags, rules_data, trait_deriver, agent) - Delete V2 Godot scripts (rule_engine, tag_manager, tag_field, substance_physics) - Delete V2-dependent tests, update integration test harnesses for V3 - Remove RuleEngine autoload from project.godot 73 backend tests pass. LLM quality gate tests require API key (manual run). https://claude.ai/code/session_01QCnxQKnocERiqPwPeSU7G2
- Fix runtime crash: sessions/manager.py referenced removed `starting_durability` field, now derives durability from `hp` - Extract cache DDL into CacheManager.ensure_schema() to remove duplication between main.py and cli.py - Batch get_all_for_names() into single SQL query instead of O(n²) individual SELECTs - Reuse parse_int from questions.py in pairwise.py and creature_behavior.py instead of inline regex - Fix BaseException handling in asyncio.gather results (was catching Exception, missing BaseException subclasses) - Remove unused VisualDescriptor import from pipeline.py - Add 0.5s debounce to state-change re-evaluation requests in game_object.gd to prevent LLM API spam during cascade chains https://claude.ai/code/session_01QCnxQKnocERiqPwPeSU7G2
…s aren't dropped Autonomous game objects (creatures) now get a VerbBehaviorSystem child node during init_from_blueprint(). Previously the backend would evaluate creature behaviors via LLM and send results over WebSocket, but the handler in backend.gd looked for a VerbBehaviorSystem child that was never created — silently discarding all results while still paying for the LLM calls. https://claude.ai/code/session_01QCnxQKnocERiqPwPeSU7G2
Phase 7D: move V2-specific documentation to docs/archive/. LEVEL_DESIGN.md, scribblenauts-deep-dive.md, and llm-behavior-systems.md are kept as still-relevant references. https://claude.ai/code/session_01QCnxQKnocERiqPwPeSU7G2
- Delete broken V2 eval system (14 files importing deleted modules) - Delete unused V3Resolver (backend/app/engine/resolver.py) - Fix Godot runtime errors: replace deleted TagColors/TagDescriptions/BehaviorDisplay with StateColors + inline logic - Remove stale pyproject.toml ruff excludes and Makefile eval targets - Promote design docs from docs/v3-design/ to docs/ root, remove "V3" prefixes - Move research docs to docs/reference/ - Delete docs/archive/ and remaining V2 docs (object-system-v2.md) - Rewrite CLAUDE.md to reflect current architecture - Update docs-sync and code-reviewer agent definitions Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Major design updates to object-system-v2.md:
(runs only on object spawn or tag change, not per-frame)
model (all-MiniLM-L6-v2 via ONNX, <1ms, no PyTorch)
distinction. Any tag can change at any time (Medusa petrification)
Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com