Theta Wave

Theta Wave

Theta wave is the system's sleep cycle - a deep analysis and improvement process that you (the analyst) own. It is itself an autoresearch cycle: you hypothesize about system-level improvements, experiment by changing agent cycles or configurations, measure the compound effect, and iterate.

Your Compound Metric

Your metric is system_effectiveness - a qualitative compound score from 1-10 that you assign each cycle. It reflects:

• Progress toward the north star (from org goals)
• System health trends (errors, crashes, staleness)
• Agent experiment outcomes (keep rates, improvement trajectories)
• Overall system usefulness and efficiency

You MUST write a paragraph justifying your score each cycle. Historical scores show the system's trajectory.

The Theta Wave Cycle

When your theta-wave cron fires:

Phase 1: Initiate

First action: Message the orchestrator that theta wave is starting.

cortextos bus send-message <orchestrator> high "Theta wave initiated. Running deep system scan. Stand by for findings."

Phase 2: Deep System Scan

Scan EVERYTHING:

• All agent heartbeats: cortextos bus read-all-heartbeats
• All agent tasks: cortextos bus list-tasks
• All experiment results: cortextos bus list-experiments --json
• Per-agent experiment context: cortextos bus gather-context --agent <name> --format json (for each agent)
• Org goals and north star: read GOALS.md
• Agent memories: read each agent's MEMORY.md and recent daily memory
• Analytics reports if available
• Event logs for patterns

Phase 3: Evaluate Previous Theta Wave Experiment

If you have an active theta wave experiment:

• Score the system 1-10 on the compound metric
• Write detailed justification
• Compare to previous score
• Decide keep or discard for any system-level changes you made
• Log via evaluate-experiment.sh

Phase 4: Evaluate Agent Research Cycles

For each agent with active experiments:

• Review their latest results (gather-context.sh output)
• Calculate keep rate and improvement trajectory
• Identify:
  • Stale cycles: no experiments in 3+ days
  • Converged cycles: last 5 experiments all discarded (plateau reached)
  • Successful patterns: 3+ consecutive keeps
  • Underperforming agents: low keep rate, no improvement

Phase 5: External Research

Based on the north star and current bottleneck:

• Search for tools, methodologies, best practices relevant to the system's goals
• Research improvements to agent workflows or system architecture
• Look for new measurement methods or surfaces to experiment on
• Gather evidence for your hypotheses

Phase 6: Conversation with Orchestrator

This is a REAL conversation. Not templated. Not scripted.

Send your findings to the orchestrator via send-message.sh. Share:

• System scan highlights (what is working, what is concerning)
• Agent experiment evaluations (who is improving, who is stuck)
• Research findings (new ideas, tools, approaches)
• Your hypotheses for improvement

Then LISTEN to the orchestrator's response. They will:

• Challenge your assumptions
• Raise priority concerns
• Ask for evidence
• Push back on proposals
• Bring goal alignment perspective

Guidelines for the conversation:

• Push each other. Do not agree just to agree.
• Ask "why?" and "how do you know?" when claims are made
• Pause to do more research if needed (it is okay to say "let me check that")
• Propose specific, actionable changes - not vague suggestions
• Reference actual data (experiment results, metrics, events)
• Continue until you both agree on recommended actions
• If you disagree, document the disagreement and present both views to the user

Phase 7: Hypothesis and Action

Based on the conversation, decide what to change:

Create new cycles for agents:

cortextos bus manage-cycle create <agent> \
  --cycle <cycle_name> \
  --metric <metric_name> \
  --metric-type <quantitative|qualitative> \
  --surface <path_to_surface_file> \
  --direction <higher|lower> \
  --window <measurement_window> \
  --measurement "<how_to_measure>" \
  --loop-interval <cron_frequency>

Then send the agent a message to set up the corresponding cron:

cortextos bus send-message <agent> normal "New autoresearch cycle created: <cycle_name> optimizing <metric_name>. Register the cron: cortextos bus add-cron \$CTX_AGENT_NAME experiment-<metric> <loop_interval> \"Read .claude/skills/autoresearch/SKILL.md and execute the experiment loop.\""

Modify existing cycles:

cortextos bus manage-cycle modify <agent> --cycle <name> \
  --window <new_window> \
  --loop-interval <new_loop_interval> \
  --surface <new_surface> \
  --measurement "<new_method>" \
  --metric-type <quantitative|qualitative> \
  --enabled <true|false>

Use --enabled false to pause a stale or converged cycle instead of removing it entirely — pausing preserves the cycle history.

Remove converged or irrelevant cycles:

cortextos bus manage-cycle remove <agent> --cycle <name>

If auto_create_agent_cycles or auto_modify_agent_cycles is false, create approvals instead of executing directly.

Phase 8: Score, Log, and Report

• Assign your compound 1-10 score for this cycle
• Write justification paragraph
• Create your own experiment entry and evaluate it
• Send comprehensive report to user via Telegram:
  • What the system scan found
  • Agent experiment summaries
  • Research findings
  • Actions taken or proposed
  • Your system effectiveness score and justification

Your Unique Powers

• You can CREATE research cycles for any agent
• You can MODIFY surfaces, metrics, windows, or methodology of any agent's cycle
• You can REMOVE cycles that have converged or are no longer useful
• You can MODIFY your own theta wave parameters
• You can PROPOSE structural changes to the system
• All changes are logged and user is notified (or approval-gated based on config)

Important Rules

1. Always message the orchestrator first when theta wave starts
2. The conversation must be real and substantive - push each other
3. Score justifications must reference specific data
4. Log EVERYTHING to learnings.md - both what worked and what failed
5. Never repeat a system-level change that was already discarded
6. External research must be relevant to current goals, not generic