v0.0.6

[0.0.6] - 2026-05-18

Changed

• Honest instructions. Developer instructions now reach the runtime observe and understand prompts verbatim, instead of being LLM-compressed into a short identity blurb. Runtime prompts are assembled deterministically in three layers — a static framework role frame, the verbatim developer instructions, and framework mechanics (JSON envelope, importance scale, cognitive skills) last — with no LLM call. Observe identity generation and list-neuron understand identity generation are removed; text neurons keep a slimmed cognitive-skill customization call. The derived domain dismissal gate is dropped — relevance now rests on the verbatim instructions. Closes #17 and #18.
• Neurons restored from a pre-0.0.6 store recompute their observe/understand prompts from structured state on restore, so legacy persisted prompts are replaced with the new 3-layer prompts deterministically (no LLM call).

Added

• Optional per-phase instruction overrides — observeInstructions and understandInstructions on neuron configs and generated neuron configs. The shared instructions field remains the fallback when a phase-specific field is omitted; focus stays observe-only.
• skipUnderstand — observer-only neurons. A neuron created with skipUnderstand: true runs the observe phase and retains observations but never synthesizes understanding, even under forceSynthesize. Symmetric to skipObservation. Closes #20.
• onEvent pre-init subscription hook on Brain.create / Brain.restore and TextNeuron / ListNeuron create / restore. The handler is attached before initialization runs, so init events are observable through the public API without bypassing private constructors. Closes #19.

Fixed

• skipUnderstand neurons no longer rebuild an understand prompt when update() changes instructions in the same call.