Conversation
…fix CI - src/ layout, package renamed core -> die (agent -> agents, evo -> neural, base_types -> types, render/plotting -> view/) - pyproject.toml + uv.lock replace unpinned requirements.txt; Python >= 3.12, numpy 2.x, torch 2.x, gymnasium 1.x; ruff config; train/video extras - Fix breakage on modern deps: np.float, cm.get_cmap, xavier_uniform, in-place multiply after Tanh that broke autograd (test_grad now backprops) - Move tests to tests/ with test_* names so pytest discovers them; new CI workflow actually installs dependencies (old one installed only pytest) - Drop dead code (data_utilities, agent/learning, scratch examples); extract research notes from core/__init__.py into docs/NOTES.md Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Simulation owns the agent-env loop; SimObserver hooks (Tqdm, Metrics) make presentation and metrics pluggable - Env: public observe(), gymnasium-style seeded reset() with an init-snapshot cache (skips costly perlin regeneration on repeat seeds) - Fix lifecycle desync (agents_die rebinding broke AgentIndexer) and phantom feeding by dead agent slots parked at (0, 0) - registry: AgentId/DynamicsId enums with tuned defaults shared by CLI, examples and tests; seedable gradient/physarum RNGs; init_params made a proper property so scripted-agent JSON save works - Tests: env physics invariants, simulation/observer contract, registry Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- FrameSource + compose_grid/to_uint8_rgb produce numpy RGB frames; Simulation.render_frames() memoizes per step across consumers - InteractivePlotter rewritten: lazy figure, interactive-backend fallback chain instead of hardcoded qtagg; InteractiveViewObserver - AnimationRecorder writes GIF/MP4 via imageio, fully headless (replaces matplotlib FuncAnimation path); Env no longer owns a renderer - minimal_run example rewritten on the new API Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Console script entry point replacing hardcoded example __main__ blocks; enum-backed agent/dynamics selection, --headless, --record, --seed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- SearcherBackend seam (generation-granular ask/tell) with evotorch 0.6.1 (pinned; semi-dormant upstream) as the first backend: PGPE/CMA-ES - EvolutionTrainer evaluates candidates through the same Simulation machinery; seeded eval modes (fixed / per-generation) - MlflowTrainLogger: full run config as params, per-generation fitness metrics, checkpoints and rollout GIFs as artifacts; sqlite backend (mlflow >= 3.14 rejects the legacy file store) - Learning smoke tests (slow): fitness improves over a tiny run, checkpoint round-trips, MLflow artifacts land Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Physarum must out-forage Brownian by a calibrated margin (measured 1.27x food consumed; asserted floor 1.15x), stable across seeds; gradient agent provably climbs chemical gradients - README: install (uv), CLI quickstart, training & MLflow usage, development workflow; milestone 3 (evolutionary Neural CA) done - gitignore experiment outputs (runs/, samples/, mlruns, *.db) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rendered channels are unbounded while imshow accepts only [0, 1] floats and logs a warning per out-of-range frame, flooding the terminal during live view. Clipping in FrameSource keeps the live view and recorded animations identical. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
`die run --view` flooded the terminal with one "Clipping input data to the valid range for imshow" warning per frame per pane, scrolling the tqdm progress bar off the screen. Root cause: the rendered channels are unbounded -- medium concentrations reach ~3.2 and agent food ~1.25 -- while imshow accepts only [0, 1] floats. It clipped them itself and logged a warning each time. Clip at the FrameSource boundary, which covers the env panes and the agent self-render alike. This is what imshow already did internally and what to_uint8_rgb already did on the recording path, so neither the live view nor the recorded animation changes: a 20-step recording is byte-identical before and after. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
AGENTS.md: architecture in one breath, the invariants and gotchas that cost real debugging time (aliasing, array identity, dead-slot masking, three-headed seeding, the coordinate seam), and workflow conventions — kept at a level that won't go stale. README: merged install into quick start, condensed learning section, development section points to AGENTS.md and docs/NOTES.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Clipping stopped imshow from warning, but the panes were still losing detail: chem1 climbs past 4.0 and pinned ~5% of the medium pane to white, the trace field settles at a steady state with 27% of the pane permanently at the top of the colormap, and env_food never exceeds ~0.5 so it sat nearly invisible next to them. Add ChannelNormalizer: a per-channel scale tracking an EMA of the channel's 99th percentile. Percentile rather than max because the fields are heavy-tailed -- a few hot pixels sit 3-4x above p99 and scaling by the max would squash everything else back into the dark. The EMA keeps brightness from pulsing as a channel drifts across a run, and a floor on the divisor stops a near-empty field from being amplified into noise. Applied where each frame is built, since only the renderer knows what a channel means. Rescaled: the three medium concentrations, the trace field (before colormapping, as the colormap itself clips), and agent_food. Left alone: the alive/alpha mask (binary, not a quantity), the medium pane's zero padding channels, and agent-owned frames such as the gradient render, which is signed data centred on 0.5 -- rescaling would move its neutral point. Measured on physarum: medium chem1 saturation 5.0% -> 1.1%, env_food distinct levels 179 -> 309, trace saturation 27% -> 1.4%. On by default, with --normalize/--no-normalize on run/replay/record. This changes recorded animations by default, and gives up cross-channel comparison: equal brightness no longer means equal concentration. Co-Authored-By: Claude Opus 4.8 <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
1 participant
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.
Summary
Full modernization of the project to a current Python + AI stack, in 7 commits (one per phase):
src/die/layout (renamed fromcore, history preserved via renames),pyproject.toml+uv.lock(Python >= 3.12, numpy 2.x, torch 2.x, gymnasium 1.x, mlflow 3.x), ruff, and a CI workflow that actually installs dependencies and discovers tests.die.sim.Simulationowns the agent-env loop with pluggableSimObserverhooks; presentation (live matplotlib view, headless GIF/MP4 recording via imageio, metrics) became observers.Envgained publicobserve()and seededreset()with an init-snapshot cache.docs/NOTES.md): evotorch 0.6.1 kept, pinned, behind a generation-granularSearcherBackendseam; the intended follow-up is a JAX env port + evosax slotting into the same seam.MlflowTrainLoggerlogs full run config, per-generation fitness metrics, checkpoints and rollout animations as artifacts (SQLite backend; mlflow >= 3.14 rejects the legacy file store).die run / record / train / replay(typer), replacing edit-the-source example scripts.Real bugs fixed along the way
np.float,cm.get_cmap,xavier_uniform— repo crashed on any modern installagents_dielifecycle rebinding desyncedAgentIndexerinit_paramsmethod-vs-property broke JSON save for scripted agentsTest plan
uv run pytest— 76 passed (unit + behavioral)uv run pytest -m slow— 2 passed (full training pipeline incl. MLflow)uv run ruff check .— cleandie run --headless,die record,die train(tiny run),die replay <ckpt>exercised end-to-endNote: commits are grouped by phase with final file contents; only the branch tip is guaranteed green, not every intermediate commit.
🤖 Generated with Claude Code