Baseline configuration profiles

Copy flags into Web Settings or .agent_runtime_prefs.json → harness.env. Secrets stay in .env only.

See Configuration basics for precedence.

Shared chats (web / desktop / TUI)

All clients read and write ~/.liminal/chats/<chatId>/ (meta.json + session.jsonl). The last active chat is stored in ~/.liminal/active_chat.json.

• AGENT_CHAT_BOOT=restore_last (default) — reopen the last active chat on process start
• AGENT_CHAT_BOOT=new_chat — always create a fresh chat when the server/sidecar/TUI starts
• AGENT_CHAT_BOOT=most_recent — open the newest chat on disk, ignoring the active pointer

Transcripts and harness context are restored from session.jsonl after restart.

Stable local dev

• AGENT_TOOL_LAZY=1
• AGENT_UI_VERBOSITY=normal
• AGENT_TOOL_BODY_ELIDE=1 and AGENT_DISTILL=1 (product defaults — stale tool bodies become pointers; huge non-code outputs distill)
• AGENT_REASONING_DEFAULT_EFFORT=medium
• AGENT_COMPLEXITY_ROUTING=1

Safety-sensitive

• AGENT_SAFETY_JUDGE=1
• AGENT_APPROVAL_TIMEOUT_MS=120000

Autonomous long-run (trusted local)

For migrations and multi-hour jobs — pair with liminal web --yolo or AGENT_YOLO=1:

• AGENT_SEND_TIMEOUT_MS=7200000
• AGENT_WORKFLOW_TIMEOUT_MS=7200000
• AGENT_CHILD_TIMEOUT_MS=1800000
• AGENT_WORKFLOWS=1
• AGENT_ALWAYS_TOOLS_PROFILE=max_autonomy
• AGENT_SELF_HEAL_LINT=1
• AGENT_YIELD_EVERY_N=8
• AGENT_SESSION_MODE=coding
• AGENT_COMPENSATION_ENABLED=1
• AGENT_MISSION_AUTONOMY=1 (chains sends while task:* stays in_progress; default on — set AGENT_MISSION_REQUIRES_YOLO=1 to require YOLO)
• AGENT_CONSOLIDATE_ON_IDLE=1 (fold session learnings on chat reset)
• AGENT_AUTO_APPROVE_TOOLS=run_lint,run_tests,git_status (optional middle ground without full YOLO)

YOLO (high-risk, temporary)

• AGENT_YOLO=1
• Prefer session-bound CLI flags:
  • npm run web -- --yolo
  • npm run web:dev -- --yolo
  • npm run tui -- --yolo

Research-heavy

• AGENT_QUERY_REWRITE=1
• AGENT_EMBED_MODEL=openai/text-embedding-3-small (or your provider’s embedding slug)
• AGENT_VAULT_AUTO_WRITE=research
• AGENT_WEB_READABILITY=1

Workflow: Harness protocol — Web research.

Long-session / deep reasoning

• AGENT_COMPRESS_SEMANTIC=1
• AGENT_REFLEXION_SEMANTIC=1 (default on)
• AGENT_PROTOCOL_INTENT_HINT=coding (adjust to task class: knowledge, execution, introspection)

Plugin extensions

• AGENT_PLUGIN_DIR=/absolute/path/to/plugins