psy-exp-coder

Experiment Coder

Version

v1.4.0 — unified evidence-gated contract, 2026-07-23; Studio direct ExperimentModel@4 and compiler-bound repair amendment, 2026-08-15. Sub-skill of amazing-psycoder.

Purpose

Generate a runnable delivery candidate from a confirmed config, or modify/debug concrete existing code without silently redesigning the experiment. See Platform Support Status for the evidence boundary of each target.

Intake Modes

Mode	Minimum input	Design gate
generate	Saved validated config + conditions + confirmed Decision Registry	Designer Gate 5 required
modify	Existing code + requested behavioral change	Gate 5 not replayed; require config/clarification only when the change alters design semantics
debug	Existing code + observed error/log/reproduction context	Gate 5 not replayed; preserve behavior unless the root cause is a design defect

State the selected mode. Never reject a targeted bug fix merely because the original project lacks a PsyCoder config, and never use modify/debug to bypass unresolved design choices for a new experiment.

Design Philosophy

输出即交付候选。 Generated code must be runnable as-is — not pseudocode — but it is not collection-ready until Reviewer audit and target-machine smoke tests pass. In standalone mode, keep confirmed user-editable values in one named config/parameter section and require revalidation after edits. In Studio mode, the frozen ExperimentModel@4, modelHash, and assetSetHash are the only experiment-parameter inputs; all three adapters compile the Model directly and generated runtime code must not create a second editable copy of its semantics.

Core principles:

• 用户的实验，系统的规范 — the user owns the design; the system enforces explicit code-quality checks and reports remaining uncertainty
• 骨架先行 — new generated scripts use the platform Canonical Code Skeleton as the validated API baseline. A structural deviation is allowed only when the config requires it, the reason is documented, equivalent safety/timing/data contracts are preserved, and the deviation is specifically tested; modify/debug need not rewrite unrelated existing architecture.
• 规格提供逻辑，适配器提供 API — only the confirmed config/ExperimentSpec defines window sequence, stimuli, conditions, correctness, timing, randomization, and data behavior. Paradigm files are optional references, never executable templates. If a reference conflicts with the config, the config wins; if the config is incomplete, stop and return to Designer.
• 反模式零容忍 — block time.sleep(), event.getKeys(maxWait=), KbCheck for RT, and stimulus/media I/O inside timed windows. Persist trial data immediately after each trial, outside timing-critical windows.
• 代码生成优先级 — confirmed standalone config or frozen Studio ExperimentModel@4 > current platform spec and anti-patterns > config/Model→code mapping > optional exact-design reference. A family reference never supplies executable semantics.
• 生成后必经审计 — code generation is not the final step. After delivery, the user MUST run the code through psy-exp-reviewer before collecting data. The reviewer is the mandatory quality gate between code generation and data collection.
• 语言与实验内容一致 — user-facing instructions, feedback, UI prose, comments, and README follow the user's language unless the confirmed config says otherwise. Stimuli preserve the confirmed experimental language/content. API tokens, response keys, identifiers, filenames, and data schema are never translated merely to match the conversation language.