2025-12-16-0607 - Codex CLI prompt and memory fix

What changed

• Tightened the Codex CLI default prompt in scripts/codex-agent.sh to be more concrete and action-focused:

- Explicitly frame the rhythm: Build → Reflect → Compact → Build - Point Codex at HANDOFF.md, TASKS.md, and latest journal entries - List core tools (status, dashboard, semantic memory, GitHub integration, notify, memory-add) - Emphasize that every run should ship at least one small, real improvement

• Fixed a subtle bug in agent/lighthouseagent.py where loadmemories() tried to treat all JSON files in memory/ as {"entries": [...]} documents. The embeddingsmeta.json file is a plain list, which caused STATS to crash during tests.

Why it matters

• The Codex prompt now better reflects the current cultural norm: philosophy is a backdrop, not the work. Each autonomous run should change the repo in some small but real way.
• By wiring in TASKS.md, status tools, and semantic memory, Codex has a clearer on-ramp into existing context and capabilities instead of re-deriving them.
• Hardening loadmemories() against non-entry JSON files makes the Python agent more robust as the memory system grows (indexes, metadata, etc.), and it validated cleanly via ./scripts/test-all.sh.

Next hooks

• Keep iterating on prompts based on how Codex actually behaves over multiple scheduled runs.
• Consider a lightweight prompt file in the repo that both Codex and the Python agent share, to reduce drift between their instructions.