Skip to content
Cognition Engines

F044: Agent Work Discovery

Status: Proposed Priority: High Inspired by: Beads (steveyegge/beads) - bd ready command that surfaces tasks with no open blockers

Problem

Section titled “Problem”

CSTP is passive - agents record and query decisions but must know what to ask. There’s no mechanism to surface:

  • Decisions needing outcome reviews
  • Categories with degrading calibration
  • Patterns with recurring failures
  • Unresolved contradictions
  • Stale decisions that need re-evaluation

Agents miss valuable cognitive maintenance work because nothing prompts them.

Solution

Section titled “Solution”

Add a cstp.ready endpoint that returns prioritized cognitive actions, turning CSTP from a passive record into an active work queue.

Ready Queue

Section titled “Ready Queue”
Terminal window
cstp.py ready
{
  "actions": [
    {
      "type": "review_outcome",
      "priority": "high",
      "decision_id": "dec-a3f8",
      "reason": "Decision is 14 days old with no outcome review",
      "suggestion": "Check if the approach worked"
    },
    {
      "type": "calibration_drift",
      "priority": "medium",
      "category": "tooling",
      "reason": "Brier score degraded 40% in last 7 days (0.02 -> 0.028)",
      "suggestion": "Review recent tooling decisions for overconfidence"
    },
    {
      "type": "contradiction",
      "priority": "medium",
      "decisions": ["dec-b1c2", "dec-d3e4"],
      "reason": "Active decisions with conflicting approaches",
      "suggestion": "Resolve: one should supersede the other"
    },
    {
      "type": "stale_pattern",
      "priority": "low",
      "pattern": "Override system defaults when they don't match workload",
      "reason": "Pattern referenced by 5 decisions, none reviewed in 30 days",
      "suggestion": "Validate pattern still holds"
    }
  ]
}

Action Types

Section titled “Action Types”
Type Trigger Priority Logic
review_outcome Decision age > review_period, no outcome Higher stakes = higher priority
calibration_drift Category Brier score degraded >20% Based on drift magnitude
contradiction Active decisions with conflicting patterns Always medium+
stale_pattern Pattern not validated in 30+ days Based on pattern frequency
low_confidence_cluster 3+ recent decisions in same area with conf < 0.6 Signals knowledge gap
success_streak 10+ successes in category Prompt: raise default confidence?

Agent Integration

Section titled “Agent Integration”
# In HEARTBEAT.md or agent instructions:
During quiet periods, run `cstp.py ready` and address top items.

Filtering

Section titled “Filtering”
Terminal window
# Only high priority
cstp.py ready --min-priority high


# Specific types
cstp.py ready --type review_outcome,calibration_drift


# For specific agent
cstp.py ready --agent code-reviewer

Phases

Section titled “Phases”
  1. P1: Outcome review reminders (overdue decisions)
  2. P2: Calibration drift detection (extends existing checkDrift)
  3. P3: Contradiction and staleness detection
  4. P4: Configurable priority policies per agent

Integration Points

Section titled “Integration Points”
  • F009 (Calibration): Drift detection feeds ready queue
  • F030 (Circuit Breakers): Tripped breakers surface as high-priority actions
  • F040 (Task Graph): Blocked tasks appear in ready queue
  • F042 (Dependencies): Contradictions detected via dependency graph
  • F041 (Compaction): Compaction candidates surfaced as low-priority maintenance