15 lines
2.2 KiB
Markdown
15 lines
2.2 KiB
Markdown
Gitea (localhost:3000): Users: rockachopa(admin), hermes(id=4), manus(id=3), kimi(id=5). Repos: rockachopa/Timmy-time-dashboard, hermes/hermes-config (private, rockachopa=admin collab). Hermes token: ~/.hermes/gitea_token. DO NOT use Alex's token — ever. Sync tool: ~/.hermes/bin/hermes-config-sync — run after modifying config/scripts/memories/skills. Future work: agentd user isolation.
|
|
§
|
|
Timmy architecture plan: Replace hardcoded _PERSONAS and TimmyOrchestrator with YAML-driven agent config (agents.yaml). One seed class, all differentiation via config. User wants to update YAML files, not Python, to add capabilities. Key files to refactor: agents/timmy.py, agents/base.py, tools_delegation/__init__.py, tools_intro/__init__.py. SubAgent class stays mostly as-is, just reads from YAML. Routing should be config-based pattern matching, not LLM calls. Per-agent model assignment (big model for orchestrator/code/research, small for simple tasks). qwen3:30b pulling as primary local model (~18GB, MoE 3B active).
|
|
§
|
|
Kimi delegation: 1-3 files max, pre-extract context (scans full codebase otherwise), ~/worktrees/ not /tmp/, verify commits, two-attempt rule. Worktrees: ~/worktrees/kimi-*.
|
|
§
|
|
2026-03-14: Issues #36-#40,#52 fixed. Voice loop, fallback chain, self-prompt queue built. Timmy on qwen3:30b (4096 ctx). Loop v2 with 20min timeout. Filed eval issues #77-#87. Status: ~/.hermes/bin/timmy-status.sh.
|
|
§
|
|
Voice loop built (timmy voice): Piper TTS + Whisper STT + Ollama, all local. Alexander wants custom voice training with his own samples eventually.
|
|
§
|
|
Hermes fallback: anthropic → local qwen3:30b only. No chain, simple single fallback. Kimi not used.
|
|
§
|
|
2026-03-15: Timmy-time-dashboard merge policy: auto-squash on CI pass. Squash-only, linear history. Pre-commit hooks (format + tests) and CI are the gates. If gates work, auto-merge is on. Never bypass hooks or merge broken builds.
|
|
§
|
|
HARD RULE: Never bypass safety mechanisms (hooks, linters, tests) for speed. Never use --no-verify. If a gate is slow, flag it — don't gut it. Default conservative. Quality gates are not obstacles. Loop prompt updated 2026-03-15: Phase 1 is now "fix broken PRs first," --no-verify banned. |