LIA

📓 The assistant's personal journal grows up

Refines the personal journal (amends ADR-079) against entries that were often useless or harmful — over-generalising a rule from a single short message, or enshrining a false "I have no access to your data" limitation that then made the assistant refuse to even try.

Highlights

• Restraint-first writing — default output is empty; an operational directive (L1) must be grounded in an explicit user signal you could quote and safe to obey blindly. Generic prohibitions forbid asserting a limit on the assistant's own capabilities/tools and generalising from a single surface feature. L0 becomes a capped, private release valve.
• Level-routed injection — operational prompts now carry only L1/L2 directives; L0 (private feedstock) and L3 (carried by the compiled portrait) are excluded by default via JOURNAL_OPERATIONAL_INJECTION_EXCLUDE_LEVELS. Extraction and consolidation still see every level.
• ReAct directive coherence — the autonomous reasoning loop now also receives L1/L2 directives (count-capped by JOURNAL_REACT_CONTEXT_MAX_ENTRIES, injected in full so a directive is never cut mid-sentence), closing a cross-mode gap.
• De-pressured consolidation — L2 synthesis only on genuine convergence ("zero L2 is the correct outcome when no convergence exists"); dedup guard against merging directives that prescribe different actions.

Notes

• No schema change, no migration.
• Tests: Ruff / Black / MyPy strict clean; 8414 backend unit tests green; i18n parity (6 languages) verified.
• See ADR-088 · CHANGELOG

🤖 Generated with Claude Code