Response Tuning

Improve LifeOS orchestrator behavior based on this feedback: $ARGUMENTS

Context

• Current branch: !git branch --show-current
• Recent commits: !git log --oneline -3
• Recent conversations: !sqlite3 data/conversations.db "SELECT m.role, substr(m.content, 1, 200), m.created_at FROM messages m JOIN conversations c ON m.conversation_id = c.id ORDER BY m.created_at DESC LIMIT 10;" 2>/dev/null || echo "NO_CONVERSATION_DB"
• Recent perf traces: !curl -s http://localhost:8000/api/perf/traces?limit=3 2>/dev/null | python3 -c "import sys,json; data=json.load(sys.stdin); [print(f'{t.get(\"question\",\"\")[:100]} | tools: {[s[\"name\"] for s in t.get(\"spans\",[]) if s[\"name\"].startswith(\"tool_\")]}') for t in data.get('traces',[])]" 2>/dev/null || echo "NO_PERF_DATA"

How this works

You are fixing the orchestrator's behavior — the instructions that tell LifeOS how to search, pick tools, integrate sources, and format responses. The user got a bad result and is telling you what went wrong so you can prevent it from happening again.

The fix is almost always a targeted edit to an instruction in a system prompt. Think of it like tuning a knob — you're adjusting the model's tendencies, not rewriting the system.

Step 1: Diagnose

Read the feedback and the recent conversation context above. Identify:

1. What went wrong — Map the feedback to a specific failure mode:

   • Didn't search enough sources → tool persistence / multi-tool patterns
   • Too verbose / too terse → response format section
   • Gave up without trying → give-up detection / self-correction
   • Wrong tool chosen → tool descriptions / when-to-use guidance
   • Poor source integration → synthesis instructions
   • Didn't follow up on partial results → search persistence rules
   • Bad Codex task execution → orchestrator scope/persistence/notification rules

2. Which file owns this behavior:

   Behavior	File	Key section
   Chat tool selection & search strategy	api/services/agent_system_prompt.py	_STATIC_PROMPT_TEMPLATE
   Chat response format & conciseness	api/services/agent_system_prompt.py	Response format section
   Chat tool round limits	api/services/agent_loop.py	max_tool_rounds param
   Give-up detection patterns	api/services/agent_loop.py	_GIVE_UP_PATTERNS
   Self-correction nudge	api/services/agent_loop.py	SELF_CORRECTION_NUDGE
   Codex task scope & persistence	api/services/claude_orchestrator.py	_SYSTEM_PROMPT
   Codex notifications	api/services/claude_orchestrator.py	NOTIFICATIONS section

3. Read the target file(s) — Read the actual current content of the file you plan to edit. Do not work from memory. The prompt may have changed since you last saw it.

Step 2: Assess scope

This skill is for small, surgical changes. Before editing, ask yourself:

• Am I changing fewer than ~20 lines of prompt text? → Proceed.

• Am I adding/adjusting a single behavioral rule or instruction? → Proceed.

• Am I tweaking a parameter (like max_tool_rounds)? → Proceed.

• Would this require restructuring the prompt, adding new tools, or changing multiple modules? → Stop. Use the /draft-issue skill to create a GitHub issue:

  Skill tool → skill: "draft-issue", args: "Orchestrator: <summary>. Feedback: <user feedback>. Root cause: <your diagnosis>. Files: <list>. Too large for a prompt tweak because <reason>."
  

  Then notify the user that an issue was created and stop.

Step 3: Edit

Make the change. Guidelines:

• Match the existing style. The prompts use a specific tone and structure — imperative, concise, example-driven. Don't introduce a different voice.
• Be specific, not vague. "Search at least 3 different sources" is better than "Be more thorough." The model responds to concrete instructions.
• Explain why when it's not obvious. A brief parenthetical like "(the user can't easily follow up from their phone)" helps the model internalize the intent.
• Don't bloat the prompt. If you're adding a new rule, check if an existing rule already covers a similar case and can be strengthened instead. Every line in the prompt costs attention — keep it lean.
• Don't weaken other behaviors. Read the surrounding context to make sure your edit doesn't contradict or undermine an existing instruction.
• Test your edit mentally. Imagine the same query that produced the bad response going through the updated prompt. Would the model now behave differently? If not, your edit isn't targeted enough.

Step 4: Branch, commit, push

# Ensure we branch from main
git checkout main
git pull --ff-only

# Create branch
git checkout -b fix/response-tuning-<short-description>

# Stage only the changed files
git add <changed-files>

# Commit
git commit -m "fix: tune orchestrator — <what was adjusted>

Feedback: <brief user feedback>
Change: <one-line description of what was edited>"

# Push
git push -u origin fix/response-tuning-<short-description>

Step 5: Notify

Report back with:

• What you diagnosed
• What you changed (quote the before/after of the edited lines)
• The branch name so the user can review

Keep it to 3-5 sentences. The user is on their phone.

Related skills

This skill delegates to other skills when needed:

• /draft-issue — When Step 2 determines the change is too large for a prompt tweak, use the Skill tool to invoke draft-issue with context about the problem and diagnosis.
• /implement — If the user later asks to implement a filed issue, point them to /implement #<issue-number>. Don't invoke it automatically from this skill.